Last year a firm handed me an AI-driven audit tool and asked me to sign off on its output. I ran it against a contract compiled with a release that had shipped eleven days earlier. The tool returned "no findings." I found a storage-collision bug in forty minutes of manual reading. The model's training corpus predated the compiler. It was not lying. It was answering a question adjacent to mine, and nobody had asked which question.
DeepMind's new genomic model โ announced as an instrument for analyzing 9 billion DNA changes โ ships in the same packaging. Enormous claim. Noble verb. No artifact I can compile.
The release is thin. I record that as a technical observation, not an insult. Per the announcement, DeepMind, a Google subsidiary, has produced an AI model positioned to analyze 9 billion DNA changes and to "democratize genetic research," with the stated aim of accelerating discovery in genetic disease understanding. That is the entire load-bearing structure. No architecture. No training corpus description. No benchmark table. No inference cost. No input specification.
I have audited ERC-20 sale contracts with more disclosure in the code comments.
Readers of this column have spent several years learning one discipline: when a protocol publishes a whitepaper and no source, you price the gap. The gap is not neutral. It is a variable, and variables get resolved by someone โ usually not the person holding the token. Genomic data is not a token, but the disclosure asymmetry is structurally identical, and the failure mode is worse. The code speaks louder than the whitepaper. DeepMind has published the whitepaper.
Take the headline number first. "9 billion DNA changes" is a scale claim, not a capability claim. gnomAD already catalogs hundreds of millions of human variants. The count of things a model has been shown is not the count of things it can classify correctly. In audit terms, that is the difference between a fuzzer's input volume and its coverage. Volume without coverage metrics is theater.
What is missing is precisely the parameter set a security reviewer would demand. Input format โ FASTA, FASTQ, or an intermediate representation? That determines whether the model sees raw reads or pre-called variants, and therefore whether it inherits the caller's errors or manufactures its own. Variant-class coverage โ SNP, indel, structural? Most published genomic models quietly excel at SNPs and degrade on structural variation, where the clinical stakes are highest. Context window โ whole genome, exome, or sliding windows? A model that handles exomes while advertising "genetic disease" is describing a subset as a whole. And the comparison set: GATK, Illumina DRAGEN, Plink. Those are the incumbents. A benchmark table against them is the minimum viable artifact. It is absent.
Then the deeper problem โ the one my 2025 white paper was about. A model that outputs pathogenic-or-benign classifications is a decision system. Decision systems must be auditable, and auditability has one hard requirement: reproducibility. If the inference path cannot be re-run deterministically โ if the weights, the tokenizer, and the data provenance are withheld โ then no regulator, no clinician, and no independent auditor can verify a single call. You are not auditing a model. You are auditing a vendor's confidence.
Trust is a vulnerability vector. That is not cynicism; it is a threat model. In lending protocols we abandoned "trusted oracle" as a design pattern because it concentrates the entire system's integrity inside one party's honesty. Genomic models repeat the concentration inside one lab's evaluation pipeline, then bolt on a clinical consequence layer.
Privacy is where the analogy breaks, and not in DeepMind's favor. An API key can be rotated. A genome cannot. DNA is the one credential in existence with no revocation path โ leaked once, it identifies the subject, their siblings, and their unconceived children, permanently. If the training corpus contains real human genomes, the differential-privacy budget, the aggregation method, and the consent chain are not optional details. They are the security perimeter. The announcement does not mention them.
Compute disclosure is likewise absent. Inference over 9 billion variant calls implies enormous TPU time. No FLOPs, no energy figure, no carbon accounting. Under the EU AI Act's high-risk classification and China's algorithmic filing regime, that omission is a compliance gap, not a formatting choice.
Here is where the bulls have a real point, and I will give it to them cleanly: AlphaFold worked. This lab's track record is not zero โ it is the best in the field. A tool that lets a researcher in Lagos run an analysis previously gated behind a five-figure pipeline is genuine value, and "democratize" is not automatically a lie.
It is, however, a pricing strategy wearing an ethics costume. Free research tiers in crypto โ public RPC endpoints, free indexers โ were never charity. They were data acquisition and ecosystem lock-in, and they worked. The same reading applies here: near-zero research access converts academic use into training signal and institutional habit, both of which monetize later through Google Cloud. That is a legitimate business. It is not a scientific claim, and the press release conflates the two.
The honest position is narrower than my readership will like. I am not saying the model fails. I am saying that as of publication you cannot know, and everyone asserting otherwise is filling the gap with brand equity.
So the ask is small and dated. Within six months: a model card, a benchmark against ClinVar and gnomAD, a variant-class breakdown, and a corpus provenance statement. Three artifacts and a table. Every serious protocol publishes less glamorous things than this before mainnet.
If nothing arrives, remember the default I work under. Every artifact is a trace of failure โ and an absent artifact is the loudest one. Volatility is just unaccounted-for variables. So is silence.
