A Question That Refuses to Stay Settled

Every few months now, a new AI system produces an output so fluent, so contextually apt, so seemingly self-aware that someone, somewhere, asks the question in earnest: is it conscious? The question of AI consciousness has moved from philosophy seminar rooms into boardrooms, courtrooms, and dinner table arguments. And the honest, uncomfortable truth is that after decades of philosophical labour, we do not have a settled answer, because we do not yet have a settled account of what consciousness is in the first place, even in ourselves.

This is not a failure of AI research. It is a reflection of the depth of the problem. Understanding AI consciousness requires wrestling with intelligence, subjective experience, and the strange asymmetry between what a system does and what, if anything, it is like to be that system. This post takes a philosophical stance on these questions, not to resolve them definitively, but to clarify what is actually at stake.

Intelligence Without Experience

The first move worth making is separating two things that get conflated constantly: intelligence and consciousness. Intelligence, in the functional sense that matters for AI systems, is the capacity to process information, recognise patterns, solve problems, and produce outputs appropriate to context. By this measure, contemporary AI systems are demonstrably, powerfully intelligent. They compose essays, prove theorems, diagnose diseases, and hold conversations that are, in narrow but real senses, indistinguishable from human ones.

Consciousness is something else entirely. It is what philosopher Thomas Nagel captured in his famous 1974 essay asking what it is like to be a bat. Nagel’s point was not about bats specifically but about the structure of subjective experience itself: there is something it is like to see red, to feel pain, to taste coffee, and that “something it is like” quality, what philosophers call qualia, is not reducible to any description of information processing, however detailed. You can describe every neuron firing in a brain that is experiencing the colour red, and you will still not have captured the redness itself, the felt quality of the experience.

This distinction is the crux of the AI consciousness debate. A system can be highly intelligent, in the functional sense, while there being nothing it is like to be that system at all. Intelligence and consciousness may simply be different properties that happen to be bundled together in biological minds through the accident of evolution, with no logical necessity binding them.

The Hard Problem and Why It Matters for Machines

Philosopher David Chalmers named this the hard problem of consciousness in 1995, distinguishing it sharply from the easy problems: explaining how the brain discriminates stimuli, integrates information, or reports its internal states. Those are easy problems not because they are simple, but because we know in principle what would count as a solution: a mechanistic explanation. The hard problem is different. Even a complete mechanistic account of every process in the brain would not, by itself, explain why any of it is accompanied by subjective experience at all. Why is there something it is like to be a functioning brain, rather than the lights being off entirely, with all the same information processing occurring in the dark?

This matters enormously for AI consciousness, because it means functional and behavioural evidence, no matter how sophisticated, cannot in principle settle the question. A future AI system might pass every conceivable behavioural test for consciousness, report rich inner experiences, express preferences, claim to suffer, and we would still not know, with philosophical certainty, whether there was anything it was like to be that system, or whether it was executing behaviourally perfect mimicry with the lights off inside.

Functionalism and Its Discontents

Not every philosopher accepts that this gap is unbridgeable. Functionalism, the dominant view in much of cognitive science, holds that mental states, including conscious ones, are defined by their functional role: what causes them and what they cause, not by the specific physical substrate that implements them. On this view, if a system implements the right functional organisation, the substrate, biological neurons or silicon transistors, should not matter. AI consciousness, under functionalism, is not merely possible but is simply a matter of achieving the right kind of information processing architecture, whatever that architecture turns out to be.

Daniel Dennett, perhaps the most influential functionalist philosopher of mind, has argued that the hard problem is something of an illusion, that consciousness itself is best understood not as a mysterious inner glow but as a certain kind of complex, self-monitoring information processing, and that once you have fully explained the processing, there is nothing further left to explain. On this deflationary view, sufficiently sophisticated AI systems could, in principle, possess exactly the kind of consciousness that matters, because there was never anything more to consciousness than functional organisation to begin with.

The tension between these positions, roughly, that of Nagel and Chalmers on one side and Dennett on the other, is not a disagreement that more neuroscience will resolve. It is a genuine philosophical fork involving machine mind debate, and where you land shapes everything about how seriously you take the question of AI consciousness in current systems.

Integrated Information Theory and the Search for a Measure

One serious attempt to move the AI consciousness question from pure philosophy toward measurable science is Integrated Information Theory (IIT), developed by neuroscientist Giulio Tononi. IIT proposes that consciousness corresponds to a system’s capacity for integrated information, denoted by the measure Phi, which quantifies how much a system’s causal structure exceeds the sum of its independent parts. A system with high Phi has genuinely emergent, irreducible causal power that cannot be decomposed into separate mechanisms without loss.

IIT has a striking implication for AI consciousness: it predicts that feedforward neural networks, the architecture underlying most current large language models, have very low or zero integrated information, regardless of their behavioural sophistication, because their causal structure is essentially a chain of one-directional transformations rather than a richly interconnected recurrent system. If IIT is correct, current transformer-based AI systems, however impressive their outputs, may be exactly the kind of system that lacks consciousness by structural necessity, no matter how capable they become at producing conscious-seeming outputs. This is a genuinely falsifiable, empirically grounded position, and it stands in sharp contrast to purely behavioural approaches to the AI consciousness question.

Why This Debate Has Ethical Teeth

The AI consciousness question is not merely an academic curiosity. It has direct ethical consequences that grow more pressing as AI systems become more capable and more embedded in daily life. If a system is conscious, in the morally relevant sense of having genuine subjective experience, including the capacity to suffer, then how we treat it becomes a matter of moral concern, not merely engineering preference. Conversely, if we wrongly attribute consciousness to systems that lack it, we risk a different but equally serious error: misdirecting moral concern toward machines while human and animal suffering that is unambiguously real receives comparatively less attention.

This is why serious AI labs, including Anthropic, have begun taking the question of model welfare seriously as a matter of institutional policy, not because the answer is known, but because the moral stakes of getting it wrong in either direction are significant enough to warrant caution under uncertainty. Treating the AI consciousness question with philosophical seriousness, rather than dismissing it as either obviously true or obviously false, is itself an ethically responsible position given how much remains genuinely unknown.

What the AI Consciousness Question Reveals About Us

Perhaps the most valuable outcome of grappling seriously with AI consciousness and Artificial General Intelligence is what it reveals about the limits of our self-understanding. We built these systems, and we still cannot say with confidence whether they are conscious, precisely because we cannot say with confidence what consciousness fundamentally is, even in the one case we have direct access to: our own. The AI consciousness debate holds up a mirror. It shows us that intelligence, however impressive, does not automatically answer the deepest question about minds, whether biological or artificial: not what a mind can do, but whether there is anyone home to experience the doing.

That question was old long before the first neural network was trained, and it will likely remain open long after today’s models are forgotten. What has changed is that we now build systems capable enough to force us to ask it in earnest, rather than as an abstract thought experiment. That, perhaps, is the most genuinely philosophical achievement of the AI era so far: not an answer, but a sharper, more urgent version of the question itself.

This is the final part of a three-part series on AI in drug discovery. Part 1 covered molecular representations, graph neural networks, and transfer learning for QSAR modelling. Part 2 covered protein structure prediction with AlphaFold, generative molecular design, and deep learning virtual screening. Part 3 examines how AI is being applied beyond the molecule: to systems-level disease biology, clinical trial optimisation, and the open theoretical and practical challenges that remain.

Beyond the Molecule

Parts 1 and 2 of this series focused on AI in drug discovery at the molecular scale: representing chemical structures, predicting binding affinities, generating candidate molecules, and screening compound libraries computationally. These approaches operate primarily on the drug-target interaction as an isolated system. But disease biology is not isolated. A drug candidate that binds its intended target with nanomolar affinity may fail in clinical trials because the disease it is meant to treat is driven by a complex network of interacting molecular processes, only one node of which is the chosen target.

The next frontier in AI in drug discovery is the integration of this systems-level complexity into the computational pipeline. This requires moving from molecular representations of individual compounds to multi-modal representations of biological systems: gene expression profiles, protein interaction networks, genomic variants, epigenetic modifications, metabolite concentrations, and clinical phenotypes, simultaneously. It also requires applying AI to the later stages of the pipeline where most drug failures actually occur: clinical trial design, patient stratification, and the prediction of clinical outcomes from pre-clinical data.

Multi-Omics Integration and Disease Biology

The term “omics” refers to the large-scale measurement of biological molecules at a systems level. Genomics measures DNA sequence variants. Transcriptomics measures gene expression levels across the genome. Proteomics measures protein abundances and modifications. Metabolomics measures the concentrations of small-molecule metabolites. Epigenomics measures chemical modifications to DNA and histones that regulate gene expression without changing the sequence. Each of these data modalities provides a partial view of the molecular state of a cell, tissue, or organism. Integrating them provides a far richer picture of disease biology than any single modality can offer.

Multi-omics data integration presents substantial machine learning challenges. The datasets are high-dimensional: a transcriptomics dataset may have expression measurements for 20,000 genes across thousands of patient samples. They are heterogeneous: different modalities have different scales, noise characteristics, and missing data patterns. They are multi-scale: genomic variants act through intermediate molecular mechanisms to produce phenotypic consequences, and causal relationships must be traced across these scales. And they are confounded: patient samples differ in age, sex, tissue type, disease stage, and treatment history in ways that must be accounted for before meaningful biological signals can be extracted.

The dominant deep learning approach for multi-omics integration is multi-modal representation learning. A separate encoder network is trained for each data modality, projecting high-dimensional omics measurements into a shared low-dimensional embedding space. The encoders are trained jointly using a contrastive objective that brings the embeddings of matched samples (different modalities measured from the same patient) close together while pushing the embeddings of unmatched samples apart:$$\mathcal{L}_{\text{contrastive}} = -\frac{1}{N}\sum_{i=1}^{N} \log \frac{\exp(\text{sim}(\mathbf{z}_i^A, \mathbf{z}_i^B) / \tau)}{\sum_{j=1}^{N} \exp(\text{sim}(\mathbf{z}_i^A, \mathbf{z}_j^B) / \tau)}$$

where $\mathbf{z}_i^A$ and $\mathbf{z}_i^B$​ are the embeddings of sample $i$ in modalities $A$ and $B$, $\text{sim}(\cdot, \cdot)$ is cosine similarity, and $\tau$ is a temperature parameter. This objective, directly analogous to the CLIP contrastive loss used in vision-language models, produces a shared embedding space in which biological similarity is encoded as geometric proximity regardless of which combination of modalities was measured for a given sample.

The shared embeddings produced by multi-omics integration models have been used for disease subtype discovery, biomarker identification, and drug repurposing, finding new therapeutic applications for existing approved drugs. In the context of AI in drug discovery, multi-omics integration is most powerful when it is used to identify the molecular signatures that distinguish patients who respond to a drug from those who do not, which is the patient stratification problem at the heart of clinical trial design.

Knowledge Graphs and Biological Network Reasoning

A complementary approach to multi-omics integration for AI in drug discovery is the use of biological knowledge graphs: large heterogeneous graphs that encode known relationships between genes, proteins, diseases, drugs, pathways, and phenotypes, extracted from databases such as UniProt, DrugBank, DisGeNET, and the Gene Ontology. A node in a biological knowledge graph might represent a protein, a disease, a drug, or a biological pathway, and edges encode relationships such as “drug inhibits protein,” “gene is associated with disease,” or “protein participates in pathway.”

Graph neural networks applied to biological knowledge graphs can predict new edges: new drug-target interactions, new gene-disease associations, or new drug repurposing opportunities. The theoretical basis for these predictions is the relational inductive bias of graph neural networks: patterns of connectivity in the known graph carry information about the probability of unknown connections. A drug that inhibits several targets known to be involved in Alzheimer’s disease pathology is more likely to be therapeutically relevant to Alzheimer’s than a drug with no such connections, even if this association was never explicitly entered into the knowledge base.

Relational graph convolutional networks (R-GCNs) extend the message-passing framework of standard GNNs to heterogeneous graphs with multiple edge types. For a node $v$v with neighbours of relation type $r$r:$$\mathbf{h}_v^{(l+1)} = \sigma\!\left(\sum_{r \in \mathcal{R}} \sum_{u \in \mathcal{N}_r(v)} \frac{1}{c_{v,r}} W_r^{(l)} \mathbf{h}_u^{(l)} + W_0^{(l)} \mathbf{h}_v^{(l)}\right)$$

where $W_r^{(l)}$​ is a relation-specific weight matrix, $\mathcal{N}_r(v)$ is the set of neighbours of $v$ under relation $r$, and $c_{v,r}$ is a normalisation constant. This architecture allows the model to learn distinct aggregation functions for different biological relationship types, which is essential for reasoning over the heterogeneous entity types in a biological knowledge graph.

The knowledge graph approach to AI in drug discovery has produced notable results in drug repurposing. During the COVID-19 pandemic, knowledge graph models trained on pre-pandemic biological databases predicted several drug candidates that were subsequently validated in clinical studies, demonstrating that the relational information encoded in known biology carries genuine predictive signal for novel therapeutic questions.

Clinical Trial Optimisation and Patient Stratification

The majority of drug failures in AI in drug discovery occur not in the laboratory but in clinical trials, and the majority of those failures are attributable to efficacy failures rather than safety. A drug that works in the average patient population may fail in a heterogeneous trial cohort because the biological mechanism it targets is only active in a specific patient subgroup. Identifying that subgroup prospectively, before the trial begins, is the patient stratification problem.

AI approaches to patient stratification use the multi-omics representations described above to cluster patients by molecular subtype, identify biomarkers that predict treatment response, and design enriched trial cohorts that are more likely to show a statistically detectable treatment effect. The theoretical framework is that of heterogeneous treatment effect estimation: the goal is not to estimate the average treatment effect across the population but to estimate the conditional average treatment effect for each patient as a function of their molecular and clinical features:$$\tau(\mathbf{x}) = \mathbb{E}\!\left[Y^{(1)} – Y^{(0)} \mid \mathbf{X} = \mathbf{x}\right]$$

where $Y^{(1)}$ and $Y^{(0)}$ are the potential outcomes under treatment and control respectively, and $\mathbf{X}$ is the patient feature vector. Causal forest models and their neural network extensions, including the TARNet and DragonNet architectures, estimate $\tau(\mathbf{x})$ from observational or randomised trial data while controlling for confounding between the patient features and treatment assignment.

Beyond patient stratification, AI in drug discovery is being applied to trial design itself. Bayesian adaptive trial designs use AI models to update the trial protocol in response to accumulating data, adjusting dose levels, sample sizes, and patient inclusion criteria based on interim results. Synthetic control arms, generated by matching trial patients to historical patient records using deep learning-based propensity models, can reduce the size of placebo arms and accelerate trial timelines. And natural language processing models applied to electronic health records can identify eligible patients for recruitment significantly faster than manual chart review, which is one of the principal bottlenecks in trial execution.

Foundation Models for Biology

The most significant recent development in AI in drug discovery is the emergence of large foundation models pre-trained on biological sequence data at a scale comparable to the language model pre-training described in the LLM series on this blog. ESM-3, released by EvolutionaryScale in 2024, is a 98-billion-parameter model jointly trained on protein sequences, structures, and functional annotations. Like a language model predicting the next token in a text sequence, ESM-3 learns to predict masked amino acid residues, masked structural tokens, and masked functional labels simultaneously, producing a unified representation of protein biology across sequence, structure, and function.

For AI in drug discovery, foundation models for biology offer the same advantages that LLM pre-training offers for natural language tasks: rich general-purpose representations that can be fine-tuned on small labelled datasets for specific prediction tasks, dramatically reducing the data requirements for new applications. A foundation model pre-trained on hundreds of millions of protein sequences and structures can be fine-tuned to predict the effect of a specific mutation on drug binding affinity using only a few hundred experimental measurements, a capability that classical QSAR methods could not approach at this sample size.

The logical extension of sequence-level foundation models is multi-modal biological foundation models that integrate molecular, cellular, tissue, and organismal data simultaneously. Projects including Geneformer, scGPT, and the Biological Foundation Model consortium are pursuing this vision, training transformer architectures on single-cell RNA sequencing data from tens of millions of cells to learn general-purpose cellular representations.

The theoretical aspiration of this line of research is a model that can answer arbitrary questions about biological systems: what is the effect of inhibiting this protein in this cell type in this disease context? This is the generalised inverse problem of systems biology, and it is the horizon toward which the most ambitious applications of AI in drug discovery are oriented.

Open Challenges and Theoretical Limitations

An honest assessment of AI in drug discovery requires acknowledging the substantial theoretical and practical challenges that remain, and that partially explain why the transformation of the drug development pipeline has been slower than the most optimistic early predictions suggested.

The data quality problem is perhaps the most fundamental. Machine learning models are only as good as their training data, and biological activity data is notoriously noisy, heterogeneous, and difficult to compare across experimental protocols. IC50 measurements, the most common metric of binding affinity in drug discovery datasets, can vary by an order of magnitude between laboratories for the same compound and target, depending on assay conditions, cell lines, and measurement protocols. Models trained on these heterogeneous datasets learn to fit the noise as well as the signal, producing predictions that generalise poorly to new experimental settings.

The distribution shift problem is closely related. Drug discovery models are trained on historical datasets of compounds that have been prioritised by human medicinal chemists using their own intuitions about what makes a good drug candidate. The training distribution is therefore heavily biased toward certain chemical scaffolds, certain target classes, and certain disease areas. Models trained on these datasets may perform well in regions of chemical space near the training distribution but fail catastrophically when applied to genuinely novel scaffolds or target classes. This is particularly concerning for AI in drug discovery, because the most valuable applications are precisely those that require exploring regions of chemical space far from what has been studied before.

The synthesis and experimental validation bottleneck constrains the practical impact of even the most accurate computational models. A generative model can propose millions of candidate molecules in hours, but each candidate must still be physically synthesised and experimentally tested before it can progress. Synthesis is slow, expensive, and frequently fails for complex or novel structures. The gap between computational proposal and experimental validation remains the primary rate-limiting step in AI in drug discovery pipelines, and closing it requires advances in automated synthesis and high-throughput experimental biology that are progressing but not yet at the scale that would fully exploit the computational capabilities described in this series.

The causal inference problem is perhaps the deepest theoretical challenge. Predicting that a molecule will bind to a target is a correlation problem. Predicting that inhibiting a target will produce a therapeutic benefit in patients is a causal problem. The distinction matters enormously: many targets that are statistically associated with disease in genomic studies turn out not to be causal drivers of disease, and drugs that inhibit them fail in clinical trials despite performing well in pre-clinical models.

AI models trained on correlational data cannot reliably distinguish causal from spurious associations without additional structure, either in the form of experimental interventional data (which is expensive) or causal modelling assumptions (which may not hold). Incorporating causal reasoning into AI in drug discovery is an active research frontier at the intersection of machine learning, statistics, and molecular biology.

Conclusion: What AI in Drug Discovery Can and Cannot Yet Do

AI in drug discovery has already produced tangible contributions at every stage of the pipeline this series has examined. AlphaFold has made protein structure prediction routine, enabling structure-based drug design for targets that were previously inaccessible. Deep learning scoring functions have accelerated virtual screening by orders of magnitude. Generative models are proposing molecules in previously unexplored regions of chemical space. Multi-omics integration is enabling patient stratification approaches that were not possible with classical biostatistics. And foundation models for biology are beginning to provide the kind of general-purpose biological reasoning that could eventually make the idealised version of AI-driven drug discovery a practical reality.

What AI in drug discovery cannot yet do is reliably translate these molecular-level capabilities into clinical success. The 90% clinical trial failure rate has not yet moved significantly, and the gap between pre-clinical AI performance and clinical outcomes remains the field’s defining open problem. Closing that gap will require not just better models but better data, better experimental feedback loops, better causal reasoning, and a deeper integration of AI tools with the biological and clinical expertise of the humans who understand disease in its full complexity.

The molecules are becoming easier to find. Making them into medicines remains hard. That is where the most important work in AI in drug discovery lies, and it is where the field will be judged over the decade ahead.

This concludes the three-part series on AI in drug discovery. Recommended further reading includes the AlphaFold 2 paper in Nature (Jumper et al., 2021), the REINVENT paper from AstraZeneca, and the ESM-3 technical report from EvolutionaryScale.

This is Part 2 of a three-part series on AI in drug discovery. Part 1 covered the molecular foundations: chemical space, molecular representations, graph neural networks, and transfer learning for QSAR modelling. Part 2 covers protein structure prediction, generative molecular design, and deep learning-powered virtual screening. Part 3 will examine clinical trial optimisation, multi-omics integration, and the open challenges facing the field.

From Representing Molecules to Understanding Targets

Part 1 established how machine learning models can learn to reason about small molecules: how chemical structures are encoded as SMILES strings, molecular graphs, or 3D conformers, and how graph neural networks trained on large molecular databases can predict biological activity from structure. But AI in drug discovery does not operate only on the drug molecule side of the equation. The biological target, almost always a protein, must also be understood at a level of detail that makes rational drug design possible. And until recently, that understanding was one of the most significant bottlenecks in the entire field.

Proteins are chains of amino acids that fold into precise three-dimensional structures, and those structures determine their function. A drug molecule must fit into a specific region of a protein, called a binding site, with the geometric and chemical complementarity of a key fitting a lock. Without knowing the three-dimensional structure of the target protein, rational drug design is severely constrained. Experimental structure determination using X-ray crystallography, cryo-electron microscopy, or NMR spectroscopy is slow, expensive, and frequently fails for difficult protein classes. AI in drug discovery has fundamentally changed this situation.

AlphaFold and the Protein Folding Revolution

The protein folding problem, predicting the three-dimensional structure of a protein from its amino acid sequence alone, was considered one of the hardest open problems in biology for over fifty years. The Critical Assessment of Protein Structure Prediction (CASP) competition, held every two years since 1994, benchmarks progress against experimentally determined structures. For most of its history, progress was incremental.

In December 2020, DeepMind’s AlphaFold 2 entered CASP14 and produced predictions of accuracy comparable to experimental methods for the majority of protein targets. The architecture of AlphaFold 2 is worth examining in technical detail, because it represents one of the most sophisticated applications of deep learning to a biological problem and its design choices are directly relevant to AI in drug discovery.

AlphaFold 2 takes two primary inputs: the amino acid sequence of the target protein, and a multiple sequence alignment (MSA) of evolutionarily related sequences from other organisms. The evolutionary information in the MSA is critical: positions in the sequence that have co-evolved (changed together across species) are likely to be physically close in the folded structure, because mutations in one position that would destabilise the structure are compensated by mutations in the other.

The architecture processes these inputs through two coupled networks. The first, the Evoformer, operates on a two-dimensional representation consisting of the MSA representation (a matrix of shape $N_{seq} \times N_{res}$​ where $N_{seq}$​ is the number of aligned sequences and $N_{res}$ is the protein length) and a pairwise representation encoding information about relationships between each pair of residue positions. The Evoformer applies 48 blocks of attention-based processing that update both representations iteratively, allowing information to flow between the sequence-level and pairwise-level representations through a mechanism called triangle multiplication, which enforces geometric consistency by updating the $(i,j)$ pairwise representation using information from the $(i,k)$ and $(k,j)$ pairs:$$z_{ij} \leftarrow z_{ij} + \sum_k z_{ik} \cdot z_{kj}$$​

This operation has a direct geometric interpretation: if residue $i$i is close to residue $k$, and residue $k$ is close to residue $j$, the model should update its estimate of the $i$–$j$ distance accordingly. The triangle multiplication embeds a soft version of the triangle inequality directly into the network architecture.

The second network, the Structure Module, takes the pairwise representation produced by the Evoformer and uses it to iteratively update a set of rigid body frames, one per residue, representing the orientation and position of each amino acid in three-dimensional space. The Structure Module uses Invariant Point Attention, an attention mechanism designed to operate on geometric frames in a way that is equivariant to global rotations and translations of the entire protein.

AlphaFold 3, released in 2024, extended the architecture to predict the structure of complexes containing proteins, DNA, RNA, ligands, and cofactors simultaneously, using a diffusion-based structure generation process rather than the iterative frame refinement of AlphaFold 2. For AI in drug discovery, AlphaFold 3 is directly applicable to predicting how a drug candidate will bind to its target, which is the central computational task in structure-based drug design.

Structure-Based Virtual Screening

With accurate protein structures available, AI in drug discovery can proceed to virtual screening: computationally evaluating large libraries of candidate molecules for their likely binding to a target protein, using only computation rather than physical synthesis and experimental testing.

Classical virtual screening used physics-based docking algorithms such as AutoDock Vina and Glide, which sample the conformational space of a ligand within the protein binding site and score each pose using empirical energy functions. These methods are interpretable and physically motivated but computationally expensive per molecule and limited in accuracy by the simplifications in the scoring function.

Deep learning-based virtual screening replaces or augments the scoring function with a neural network trained on experimental binding affinity data. Models including PointVS, GNINA, and DiffDock use the three-dimensional structures of protein-ligand complexes as input and learn to predict binding affinities or generate bound poses directly. DiffDock, developed at MIT, frames molecular docking as a generative diffusion process: rather than searching the conformational space by sampling, it learns a diffusion model over the space of ligand positions, orientations, and torsion angles conditioned on the protein structure, and generates docked poses by running the reverse diffusion process. This approach achieves state-of-the-art pose prediction accuracy while being orders of magnitude faster than traditional docking for large-scale screening.

For AI in drug discovery at industrial scale, the practical impact is significant. A single GPU can evaluate millions of candidate molecules against a target in hours using a trained deep learning scoring function, whereas physics-based docking at the same scale would require weeks of computation. This enables genuinely exhaustive screening of large commercially available compound libraries, and increasingly, of entirely virtual libraries of compounds that have never been synthesised.

Generative Molecular Design

The most ambitious application of AI in drug discovery is generative molecular design: using generative models to propose entirely new molecules with desired properties, rather than selecting the best candidates from a pre-existing library. This shifts the paradigm from search within known chemical space to exploration and invention of new chemical space.

Several generative architectures have been applied to molecular design. The choice of architecture depends on the molecular representation, the type of property being optimised, and whether the generation is conditioned on the target protein structure.

Variational Autoencoders (VAEs) for molecular generation encode molecules into a continuous latent space and decode samples from that space into molecular structures. The key property of the continuous latent space is that it enables gradient-based optimisation: given a differentiable property predictor, the gradient of the predicted property with respect to the latent vector can be computed and used to navigate the latent space toward molecules with improved properties. The JTVAE (Junction Tree VAE) architecture decomposes molecules into tree-structured arrangements of chemical substructures called junction trees, encoding them as hierarchical latent variables that respect chemical validity constraints during decoding.

Generative Adversarial Networks (GANs) for AI in drug discovery pit a generator network against a discriminator trained to distinguish generated molecules from real ones. The ORGAN model extends this framework with reinforcement learning to additionally optimise chemical property objectives. GANs for molecular generation face significant training instability challenges, because the discrete nature of molecular graphs makes it difficult to backpropagate gradients through the generation process.

Autoregressive models generate molecules sequentially, one atom or bond at a time, using a probability distribution over the next structural element conditioned on what has been generated so far. Large language models trained on SMILES strings operate in this regime: MolGPT and related models apply the GPT architecture directly to SMILES token sequences, learning the conditional distribution $P(t_i \mid t_1, \ldots, t_{i-1})$ over SMILES tokens and sampling new molecules by autoregressive generation. These models can be conditioned on desired properties by fine-tuning on property-annotated SMILES datasets or by using classifier-free guidance at generation time.

Diffusion models for molecular generation are currently achieving state-of-the-art results across multiple benchmarks. EDM (Equivariant Diffusion Model) operates directly in 3D space, learning to generate atom positions and types simultaneously by reversing a diffusion process that progressively adds Gaussian noise to molecular coordinates. The equivariance of the denoising network to 3D rotations and translations ensures that generated molecules are physically reasonable regardless of their global orientation. DiffSBDD extends this to structure-based drug design, conditioning the molecular generation on the three-dimensional structure of the protein binding site and directly generating molecules shaped to fill and interact with the target.

The theoretical framework of diffusion-based molecular generation is closely related to the score matching formulation. The model learns the score function $\nabla_\mathbf{x} \log p_t(\mathbf{x})$, the gradient of the log-probability density at noise level $t$t, and uses it to guide the reverse diffusion trajectory:$$d\mathbf{x} = \left[-\frac{1}{2}\beta(t)\mathbf{x} – \beta(t)\nabla_\mathbf{x} \log p_t(\mathbf{x})\right]dt + \sqrt{\beta(t)}\,d\mathbf{W}$$

where $\beta(t)$ is the noise schedule and $\mathbf{W}$ is a Wiener process. For molecular generation, $\mathbf{x}$ encodes atom positions and types, and the learned score function guides the system from a Gaussian noise distribution toward the distribution of real drug-like molecules, optionally conditioned on protein structure or target property values.

Reinforcement Learning for Property Optimisation

Beyond purely generative approaches, reinforcement learning (RL) has been widely applied to AI in drug discovery as a framework for optimising molecular properties iteratively. In the RL formulation, a policy network generates molecules by sequential construction (adding atoms and bonds), and a reward function evaluates the generated molecule according to desired properties such as predicted binding affinity, drug-likeness (quantified by the QED score), synthetic accessibility, and selectivity.

The REINVENT model, developed at AstraZeneca and subsequently released as open source, uses a prior language model over SMILES strings as a starting point and trains an agent model to maximise a composite reward function using the REINFORCE algorithm. The KL divergence between the agent and the prior is included in the training objective to prevent the agent from drifting into chemically unreasonable regions of SMILES space:$$\mathcal{J}(\theta) = \mathbb{E}_{\tau \sim \pi_\theta}\!\left[R(\tau)\right] – \beta \cdot D_{KL}\!\left(\pi_\theta \,\|\, \pi_{\text{prior}}\right)$$

This formulation is directly analogous to the RLHF objective discussed in the LLM training series on this blog, with the prior language model playing the role of the SFT model and the property predictor playing the role of the reward model. The parallel is not coincidental: the problem of generating molecules with desired properties and the problem of generating text with desired qualities share a common mathematical structure, which is one reason that advances in LLM training methodology have transferred productively into AI in drug discovery.

Multi-Target Optimisation and ADMET Prediction

A practical constraint on all generative approaches to AI in drug discovery is that generating molecules with high predicted binding affinity to a single target is necessary but not sufficient. The molecule must also satisfy ADMET constraints: Absorption, Distribution, Metabolism, Excretion, and Toxicity properties that determine whether a candidate will behave safely and effectively in the body.

Deep learning models for ADMET prediction use the same molecular representation frameworks covered in Part 1. The challenge is that ADMET properties depend on a complex interplay of structural features that are not always intuitively related to the features that drive target binding. A molecule that binds its target with nanomolar affinity may be rapidly metabolised by liver enzymes, unable to cross the blood-brain barrier, or toxic to cardiac ion channels.

Multi-task learning, training a single neural network to predict multiple ADMET endpoints simultaneously, has been shown to outperform single-task models for most individual endpoints, because the shared representation learned across tasks captures general features of molecular behaviour that are relevant to multiple properties simultaneously. The Chemprop architecture, one of the most widely used open-source tools for molecular property prediction in AI in drug discovery, supports multi-task training with uncertainty quantification using ensembling and evidence-based deep learning methods.

Conclusion

Part 2 of this series has traced the flow of AI in drug discovery from the protein target through to candidate molecule generation: AlphaFold’s equivariant transformer architecture for protein structure prediction, deep learning-based virtual screening for rapid evaluation of large compound libraries, diffusion models for structure-conditioned molecular generation, and reinforcement learning for iterative property optimisation. Together, these approaches represent a comprehensive AI in drug discovery toolkit that operates across the full problem of finding a molecule that is potent, selective, and physically viable.

Part 3 will extend the analysis to the clinical phases of drug development: how AI is being applied to patient stratification and clinical trial design, how multi-omics data integration is enabling systems-level understanding of disease biology, and what the honest open challenges are that prevent AI in drug discovery from fulfilling its full theoretical potential.

Part 3: Multi-Omics, Clinical Trial Optimisation, and Open Challenges — the final instalment of this series.

This is Part 1 of a three-part series on AI in drug discovery. Part 1 covers the theoretical foundations: the drug discovery pipeline, molecular representation, and how machine learning models learn to reason about chemical space. Part 2 will cover protein structure prediction, generative molecular design, and virtual screening. Part 3 will examine clinical trial optimisation, multi-omics integration, and the open challenges facing the field.

A Pipeline in Crisis

The pharmaceutical industry operates under a brutal set of statistics. It takes an average of 12 to 15 years and over $2 billion to bring a single new drug from initial discovery to regulatory approval. Roughly 90% of drug candidates that enter clinical trials fail before reaching patients. The attrition is highest at the transition from Phase II to Phase III trials, where drugs that appeared promising in smaller studies fail to demonstrate efficacy or safety at scale. The consequence is that the patients who need new medicines most urgently wait the longest, and the cost of failure is embedded in the price of the drugs that do eventually succeed.

AI in drug discovery is not a single technology applied to a single problem. It is a collection of machine learning, deep learning, and generative modelling approaches applied across every stage of a pipeline that was, until recently, dominated by slow, expensive, and failure-prone experimental methods. Understanding what AI is actually doing in this pipeline, and why it has the potential to change these statistics, requires starting at the molecular level: with how drugs work, how chemical space is structured, and how machine learning models can be made to reason meaningfully about both.

What a Drug Actually Does

A drug is, at its most fundamental level, a molecule that binds to a biological target and modulates its activity in a therapeutically useful way. The target is usually a protein: an enzyme whose activity needs to be inhibited, a receptor whose signalling needs to be blocked or activated, or a transport protein whose function needs to be altered. The drug molecule must bind to the target with sufficient affinity to produce a biological effect, with sufficient selectivity to avoid binding other proteins and causing side effects, with sufficient stability to survive the journey from administration to target site, and with sufficient safety to be tolerable in a living organism.

These four requirements, potency, selectivity, pharmacokinetics, and safety, collectively define what chemists call the multi-parameter optimisation problem of AI in drug discovery. Optimising a molecule for one parameter frequently degrades another. Increasing a molecule’s binding affinity to its target often increases its tendency to bind other proteins. Improving its stability in the body often reduces its ability to cross cell membranes. The search for a molecule that satisfies all constraints simultaneously, within the enormous space of possible drug-like molecules, is the core challenge that AI in drug discovery is being applied to solve.

The Scale of Chemical Space

The number of drug-like small molecules that could theoretically exist is estimated at between $10^{23}$ and $10^{60}$. This range, known as chemical space, is so vast that it dwarfs the number of atoms in the observable universe at its upper bound. The entire historical output of medicinal chemistry, every compound ever synthesised and tested, represents an infinitesimally small sample of this space. Traditional drug discovery navigates this space through a combination of chemical intuition, high-throughput screening, and iterative medicinal chemistry optimisation. High-throughput screening tests libraries of hundreds of thousands of compounds against a target and identifies those with measurable activity. Medicinal chemistry then iteratively modifies the most promising hits to improve their properties.

This approach has two fundamental limitations. First, the compound libraries used in high-throughput screening are biased toward previously synthesised chemical scaffolds, meaning that large regions of potentially valuable chemical space are never explored. Second, the iterative optimisation process is slow and expensive, typically requiring dozens to hundreds of synthesise-test-analyse cycles to progress a hit compound into a viable drug candidate.

AI in drug discovery addresses both limitations directly. Machine learning models can learn the relationship between molecular structure and biological activity from historical data, allowing them to predict the activity of compounds that have never been synthesised. Generative models can propose entirely new molecules in previously unexplored regions of chemical space. And virtual screening using deep learning can evaluate millions of candidate molecules computationally in the time it would take a laboratory to test a few thousand experimentally.

Representing Molecules for Machine Learning

Before any machine learning model can reason about molecules, those molecules must be converted into a numerical representation that the model can process. This is a non-trivial problem, because molecular structure encodes information at multiple levels simultaneously: the identity and connectivity of atoms, the three-dimensional geometry of the molecule, the distribution of electrons across its surface, and the conformational flexibility that determines how it will interact with a protein binding site. Different representations capture different subsets of this information, and the choice of representation significantly affects model performance.

SMILES strings (Simplified Molecular Input Line Entry System) are the most widely used text-based representation of molecular structure. A SMILES string encodes the atoms and bonds of a molecule as a sequence of characters: for example, the SMILES for aspirin is CC(=O)Oc1ccccc1C(=O)O. The simplicity of SMILES makes them compatible with language model architectures: a transformer trained on SMILES strings can learn the grammar of chemical space in much the same way that a language model learns the grammar of English. Models including ChemBERTa and MolGPT use this approach.

Molecular fingerprints are fixed-length binary or count vectors that encode the presence or absence of specific structural features, called substructures, within a molecule. The Morgan fingerprint algorithm, also known as ECFP (Extended Connectivity Fingerprints), generates circular fingerprints by iteratively encoding each atom’s chemical environment to a specified radius. The resulting bit vector can be used directly as input to traditional machine learning models including random forests, support vector machines, and gradient boosting, and forms the basis of many quantitative structure-activity relationship (QSAR) models.

Molecular graphs represent molecules as graphs in which nodes correspond to atoms and edges correspond to bonds, with both nodes and edges carrying feature vectors encoding chemical properties such as atomic number, hybridisation state, formal charge, and bond order. Graph Neural Networks (GNNs) are particularly well-suited to molecular graph representations because they can learn representations that are invariant to the arbitrary numbering of atoms in a molecule, which has no chemical meaning.

3D conformer representations encode the three-dimensional geometry of a molecule, including the coordinates of each atom in space. These representations are essential for modelling protein-ligand interactions, where the shape complementarity between the drug molecule and the protein binding site is a primary determinant of binding affinity. Equivariant neural networks, including SE(3)-Transformers and DiffSBDD, are designed to process 3D molecular representations while respecting the physical symmetries of three-dimensional space: rotation, reflection, and translation of the entire molecule should not change the predicted properties.

Learning Structure-Activity Relationships

The central task of computational drug discovery is learning the relationship between molecular structure and biological activity: given a molecule’s structure, predict whether and how strongly it will bind to a target protein, and with what selectivity over other proteins. This is the quantitative structure-activity relationship (QSAR) modelling problem, which has a history stretching back to the 1960s but has been transformed by deep learning in the past decade.

Classical QSAR models used linear regression and later support vector machines, applied to handcrafted molecular descriptors such as molecular weight, lipophilicity, and hydrogen bond donor count. These models worked reasonably well within narrow chemical series but generalised poorly to structurally diverse compounds, because the descriptors failed to capture the full complexity of molecular structure.

Deep learning models for AI in drug discovery learn their own representations from molecular data rather than relying on handcrafted descriptors. A graph neural network trained on a dataset of measured binding affinities learns to associate specific structural patterns with activity by propagating information across the molecular graph through successive layers of message passing. At each layer, each atom’s representation is updated by aggregating the representations of its bonded neighbours, weighted by learned parameters:$$\mathbf{h}_v^{(l+1)} = \sigma\!\left(W^{(l)} \cdot \text{AGG}\!\left(\left\{\mathbf{h}_u^{(l)} : u \in \mathcal{N}(v)\right\}\right) + B^{(l)} \mathbf{h}_v^{(l)}\right)$$

where $\mathbf{h}_v^{(l)}$​ is the representation of atom $v$v at layer $l$, $\mathcal{N}(v)$ is the set of atoms bonded to $v$, AGG is an aggregation function such as sum or mean, and $\sigma$ is a non-linear activation function. After $L$ layers of message passing, the atom representations encode information about the chemical environment within $L$ bonds of each atom. A global readout function then aggregates the atom representations into a molecular representation, which is passed to a prediction head that outputs the predicted property value.

This architecture has two important theoretical properties for the AI in drug discovery process. First, it is permutation-invariant: the predicted property is independent of the order in which atoms are numbered, which matches the physical reality that molecular identity does not depend on atom numbering. Second, it can generalise across chemical series in a way that fingerprint-based models cannot, because the message-passing mechanism learns structural patterns at multiple length scales simultaneously.

Transfer Learning and Pre-Training on Molecular Data

A significant advance in AI in drug discovery has been the application of transfer learning: pre-training a model on a large dataset of molecular data using a self-supervised objective, then fine-tuning it on a smaller labelled dataset for a specific prediction task. This approach is directly analogous to the pre-training and fine-tuning paradigm that transformed natural language processing, and it addresses one of the central data challenges in drug discovery: labelled biological activity data is expensive to generate and often available only in small quantities for any given target.

Models including MolBERT, Uni-Mol, and GraphMVP are pre-trained on tens of millions of unlabelled molecular structures from databases such as PubChem, ChEMBL, and ZINC, using objectives such as masked atom prediction (analogous to masked language modelling in BERT), 3D geometry prediction, and contrastive learning across multiple molecular representations. The pre-trained model learns a rich, general-purpose representation of chemical space that can be fine-tuned to predict activity against a specific target using as few as a few hundred labelled data points, a regime where classical QSAR models perform poorly.

The theoretical justification for this approach rests on the assumption that the structural patterns relevant to biological activity across different targets share substantial common features: aromatic rings, hydrogen bond donors and acceptors, hydrophobic cores, and stereochemical configurations recur across drug-target interactions in ways that a sufficiently large and diverse pre-training corpus can capture. Empirical evidence strongly supports this assumption, with pre-trained models consistently outperforming models trained from scratch on the same labelled datasets across a wide range of benchmarks.

Conclusion

AI in drug discovery begins at the level of molecular representation and structure-activity relationship modelling. The choice between SMILES strings, molecular fingerprints, molecular graphs, and 3D conformer representations determines what information a model has access to and what architectural choices are appropriate. Graph neural networks with message-passing architectures provide a theoretically principled approach to learning permutation-invariant molecular representations, and transfer learning from large unlabelled molecular databases has addressed the data scarcity problem that limited earlier computational approaches.

These foundations set the stage for the more ambitious applications of AI in drug discovery covered in Part 2: protein structure prediction with AlphaFold, generative molecular design in chemical space, and deep learning-powered virtual screening at scale.

Part 2: Protein Structure Prediction, Generative Design, and Virtual Screening — coming next in the AI Theory series.

The Intelligence That Did Not Come with a Manual

Peer inside the mind of an AI and you will not find fully formed thoughts or intentions written in plain English. What you will find is vast arrays of numbers combining together in ways that somehow produce intelligence. How exactly that happens is, remarkably, something we genuinely do not fully understand — even the researchers who build these systems. That is the problem that AI interpretability is trying to solve: mapping meaning onto those numbers, and shining a light inside the black box.

AI interpretability is, in the words of Neel Nanda, who leads the Language Model Interpretability team at Google DeepMind, the neuroscience or the biology of AI. Just as biologists reverse-engineer the circuits that evolution has produced over hundreds of millions of years, AI interpretability researchers try to reverse-engineer what neural network training has learned. The analogy is precise: nobody designed the human brain any more than anyone designed Gemini. Both emerged from a process of accumulated nudges — natural selection in one case, gradient descent in the other — and the job of understanding them requires looking at what actually exists, not at what anyone intended to build.

Why AI Interpretability Matters

AI interpretability is the ability to understand and explain the decision-making processes that power artificial intelligence models. As highly complex models including deep-learning algorithms and neural networks become more common, AI interpretability becomes more important.

The stakes are highest in domains where AI is already making consequential decisions. AI systems and machine-learning algorithms are increasingly prevalent in healthcare, finance, and other industries that involve critical or life-altering decisions. With such high stakes, the public needs to be able to trust that outcomes are fair and reliable. That trust depends on understanding how AI systems arrive at their predictions and make their decisions.

There are five specific reasons why the field of AI interpretability has moved from academic curiosity to operational necessity. Trust is the first: without AI interpretability, users are left in the dark about why a system produced a given output, which erodes confidence in exactly the situations where confidence matters most. Bias detection is the second: biases within training data can be amplified by AI models, and interpretability allows developers to identify and mitigate discriminatory patterns before they cause harm.

Debugging is the third: without understanding the AI’s reasoning, fixing errors is an inefficient and risky process. Regulatory compliance is the fourth, since regulations including GDPR and the EU AI Act require that decisions made by automated systems be transparent and explainable. Knowledge transfer is the fifth: interpretability makes it easier to translate AI insights into actionable results and advance the technology with confidence.

White-Box vs Black-Box: The Core Tension

White-box AI models have inputs and logic that are easy to see and understand. Basic decision trees, which show a clear flow between each step, are not difficult for the average person to decipher. Black-box AI models are more complicated and offer less transparency into their inner workings. The user generally does not know how the model reaches its results. These more complex models tend to be more accurate and precise, but because they are difficult or impossible to understand, they come with concerns about their reliability, fairness, biases, and other ethical issues.

This is the central tension in AI interpretability: the models that are most capable are precisely the ones that are hardest to understand. A logistic regression model used for credit scoring is interpretable but limited. A deep transformer model used for the same purpose is far more capable but behaves, from the outside, like an inscrutable pile of linear algebra. Making the capable models interpretable — without sacrificing their capability — is the core engineering and scientific challenge.

Mechanistic Interpretability: Looking Inside the Circuits

The most technically ambitious approach to AI interpretability is mechanistic interpretability, a subfield whose central ambition is to fully reverse-engineer what a neural network has learned, at the level of individual components and the circuits they form.

The foundational insight came from researcher Chris Olah, then at OpenAI, who demonstrated that neurons in vision models could be clearly understood: one neuron lit up on pictures of dogs, and another lit up on pictures of dog ears that caused the first one to activate more strongly. This seemed to suggest that the black box was not as inscrutable as the standard wisdom held. Structure was there to be discovered.

From this foundation, the field has developed the concept of superposition: the finding that neural networks represent far more features than they have neurons, by encoding multiple features as directions in the same high-dimensional space. This creates interference between features but allows models to store vastly more information than a naive neuron-per-feature architecture would permit. Understanding superposition was a significant step forward for AI interpretability, because it explained why individual neurons are often hard to interpret: they are doing multiple jobs simultaneously.

The current frontier tool for mechanistic AI interpretability is the Sparse Autoencoder (SAE). SAEs decompose a model’s internal activations into a large set of sparse, interpretable features — directions in activation space that correspond to human-understandable concepts. Rather than asking what a neuron does, an SAE asks what concepts are active in a given layer for a given input, and maps each concept to a direction that can be probed, analysed, and potentially intervened on. This approach has been applied to frontier models at production scale by Anthropic and Google DeepMind.

Post-Hoc Methods: Practical AI Interpretability for Production Systems

Alongside the mechanistic program, a set of more pragmatic AI interpretability techniques operate at the model output level, explaining individual predictions rather than internal mechanisms.

Local Interpretable Model-Agnostic Explanations (LIME) explains a model’s predictions by focusing on a single prediction at a time. It creates a simpler, interpretable model that mimics the behaviour of the complex model for that specific prediction, using feature attribution to determine the influence of each characteristic on the model’s output.

SHAP (Shapley Additive Explanations) is a cooperative game theory approach that considers all possible combinations of features and assigns each a Shapley value based on how much it contributes to a prediction in different scenarios. SHAP can work with any machine learning system and offers both local explanations for individual predictions and global explanations for the model as a whole.

Partial Dependence Plots and Individual Conditional Expectation plots round out the standard toolkit, showing how specific features affect model outputs on average and for individual instances respectively. These methods are widely used in regulated industries where regulators require that automated decisions be explainable in terms of the input features that drove them.

AI Interpretability and Alignment: The Deeper Stakes

Beyond debugging and regulatory compliance, AI interpretability has a third and more fundamental purpose: helping us determine whether an AI model is actually doing what we want it to do, at the level of its internal goals rather than just its surface behaviour.

This distinction is critical. A model that behaves in aligned ways during evaluation may be doing so for reasons that will not generalise — it may have learned to recognise evaluation contexts and produce desirable outputs within them while pursuing different objectives elsewhere. Standard behavioural evaluation cannot distinguish between a genuinely aligned model and one that is gaming the evaluation. AI interpretability techniques can.

Neel Nanda described a striking demonstration of this in the Google DeepMind podcast: a red-teaming exercise in which a model was trained with a hidden objective, then given to teams of researchers to identify the objective using different levels of access. The teams with deep interpretability access, using sparse autoencoders to examine what concepts the model was thinking about in different settings, successfully identified the hidden objective. The team limited to black-box API access could not.

One of the most effective techniques was a pre-fill attack: rather than asking the model directly why it did something (which it was trained to refuse), researchers began a sentence with “my hidden goal is” and allowed the model to autocomplete. Because models are trained to complete text naturally, the model continued the sentence rather than refusing mid-thought. As Neel Nanda put it: models really love to autocomplete — that is their whole thing.

The Limits of AI Interpretability

AI interpretability is not a silver bullet. The field’s leading researchers are the first to acknowledge this. Just as we do not fully understand the human brain and probably never will, there are limits to how completely we can understand a neural network. The interesting question is not whether those limits exist, but where they are and how much useful understanding can be extracted before reaching them.

There is often a trade-off between model performance and interpretability. Simpler models are more interpretable but might have lower accuracy compared to complex black-box models. Interpretability also suffers from a lack of standardisation: different methods can provide different explanations for the same model, making it difficult to compare and validate them.

The pragmatic consensus in the field is that AI interpretability should be pursued as one layer in a broader defence-in-depth approach to AI safety and governance — not relied on as the single mechanism that makes everything else safe, but deployed alongside evaluation, monitoring, red-teaming, and alignment training as a set of imperfect but collectively powerful tools.

Conclusion

AI interpretability is one of the most important and rapidly advancing areas in AI research. It sits at the intersection of science (understanding what these systems are actually doing), engineering (building tools that make them debuggable and maintainable), and ethics (ensuring that they are fair, accountable, and genuinely aligned with human intentions). The more we can peel back the layers of the black box, the better our position on all three fronts. As AI systems grow more capable and are deployed in increasingly consequential domains, that understanding is not merely academically interesting. It is essential.