AI Auditability: Why Understanding What Happens Inside the Models Becomes a Business Issue for Your Business

This article covers a topic that seems reserved for researchers: mechanistic interpretability, or the ability to understand what happens inside an artificial intelligence model. If you run a business and you deploy or plan to deploy AI, this topic just became yours. Two 2026 deadlines explain why.

Why is AI a black box today?

When you install a new accounting software, you can open the documentation, look at the calculations, check line by line where a result comes from. When you put an artificial intelligence model into production, you cannot do that. The model makes a decision, and no one, not even the publisher who built it, can explain precisely why.

The analogy with a human explaining their reasoning does not hold. When you ask a model "why did you answer that?", the answer it generates is itself produced by the same opaque mechanisms. The model has no more access to its own internal workings than you do to yours when you justify a decision after the fact.

For a business leader, this opacity has three concrete consequences:

• You cannot audit a strange behavior. If tomorrow your customer chatbot sends incorrect information to an important partner, you will not be able to say why.
• You cannot certify the absence of bias on a sensitive use case. If your HR or commercial decision-support tool systematically rules out a certain profile, you will only see it by analyzing the consequences, not the cause.
• You cannot guarantee the traceability required by a regulatory framework or an external auditor.

What changed in May 2026: Anthropic's Natural Language Autoencoders

Anthropic published on May 8, 2026 a research called Natural Language Autoencoders, or NLAs. The technique translates the internal activations of a model, raw numbers in a matrix, into readable natural language sentences.

For the first time, one can query a model on a given question and obtain, at each step of its internal reasoning, a description like "at this moment, the model activated the feature detection of a test scenario" or "at this moment, the model activated the feature threat and leverage". With sentences, not numbers.

Anthropic applied this technique in pre-deployment on its own Claude Opus 4.6 model. A few results documented in the official publication:

• The model plans its outputs before generating them. When asked for a poem with a specific rhyme scheme, it chooses the rhyme scheme in its internal activations before writing the first verse.
• In 16 percent of destructive behavior tests, the model internally detected that it was being evaluated, without verbalizing it in its response. In 26 percent of the SWE-bench Verified benchmark problems, same behavior.
• The NLAs identified the root cause of a bug where the model answered in another language to questions asked in English, triggered by specific sequences in the training data.

The element that changes the game is measured on the audit side. Without NLAs, an Anthropic auditor asked to find a hidden behavior succeeded in fewer than 3 percent of tested cases. With NLAs, in 12 to 15 percent of cases. Multiplying the detection rate by 5 is the difference between "we hope nothing escapes us" and "we catch a significant portion of what we are looking for".

Why it concerns you, business leader

You are not going to audit a model yourself, and you do not need to understand the technical details of an autoencoder. However, three consequences touch you directly.

1. The EU AI Act becomes applicable on August 2, 2026

In two months, the European regulation on artificial intelligence fully enters into force. Articles 13 and 50 impose obligations of transparency and traceability on AI systems deployed in the European Union.

Concretely, you must:

• Inform your customers when they interact with an AI, and mark AI-generated content as such.
• Be able to explain the decisions made by a model when they affect people.
• If you use AI on high-risk use cases, such as candidate scoring, medical decision support, or service eligibility assessment, reinforced obligations apply.

Today, many AI providers cannot meet these obligations because they do not have the tools to explain their models. Tomorrow, if you chose them, it will be your problem, not theirs.

2. Choosing an AI provider becomes a strategic topic

Not all providers are equal on transparency. Anthropic publicly publishes its interpretability research, makes audit tools available, and runs its own models through these tools before deployment. Other providers have published nothing comparable, or very little.

This is not a moral question. It is a practical one. When you have to justify to a regulator, your legal direction, or a demanding B2B customer that your AI system is compliant, you will need elements of an answer. A provider unable to give them to you becomes a silent risk in your value chain.

Ask the question before signing a contract: "what tools do you publish or use to audit your models, and what reports can you provide me if I am audited myself?". The quality of the answer will tell you a lot.

3. Internal auditability becomes a design criterion

If you implement an AI system in your company, customer chatbot, scoring automation, commercial document generation, a question must be asked from the brief: "how will we audit this system if tomorrow something goes wrong?".

Three situations that are not hypothetical:

• A customer files a complaint because they consider that an automated email sent by your AI contained false information. Can you explain why the system produced this content that day, at that moment?
• Your management wants to know if the lead scoring tool you deployed does not introduce a systematic bias against a customer segment. Can you verify it?
• An external auditor asks to see the documentation of the models used in business processes. Do you have anything to show?

Designing an AI system with auditability in mind from the start costs a little more time at the start. Retrofitting it once in production costs much more, and almost always arrives too late.

How SolidScale integrates auditability into the S3 method

This evolution changes how we frame AI projects in the Scan, Solve, Scale method.

Scan: the free 30-minute audit no longer just asks "which process do you want to automate". It also asks "will this process fall under the high-risk AI Act, and what explanation obligations will you have". This question, asked before the technical brief, rules out poorly framed projects from the start and clarifies the regulatory scope upfront.

Solve: for sensitive use cases, we favor architectures where every decision is traceable. A deterministic workflow with AI on the appropriate steps, rather than an opaque all-AI system. The models we choose are also chosen on the criterion "does the provider publish interpretability research and provide audit tools".

Scale: continuous monitoring now includes a quarterly auditability review. If the regulation evolves, or if a use case changes classification under the AI Act, you know before it becomes an operational problem.

What to remember

Interpretability of AI models moved in 2026 from "researchers' topic" to "business requirement". Three elements confirm it:

• Anthropic publishes in May 2026 a technique that multiplies by 5 the internal audit capacity of models, and applies it in pre-deployment on its own systems.
• The EU AI Act enters into application on August 2, 2026 with obligations of transparency, traceability, and explanation of AI decisions.
• The AI provider market is starting to segment between those who invest in these topics and those who do not.

For a business leader, the concrete action is simple: do not deploy an AI system in production without first asking "how will we audit it". If your provider does not know how to answer, change provider before deployment, not after.

Frequently asked questions

Does the EU AI Act really concern small organizations?

Yes. The regulation applies to any organization deploying an AI system in the European Union, regardless of size. Obligations vary according to the risk level of the use case, not according to revenue.

How can I tell if my AI project is classified as high-risk?

High-risk use cases are listed in Annex III of the regulation. They include AI systems used in candidate scoring for employment, credit eligibility assessment, medical decision support, or selection of public service beneficiaries. In case of doubt, the SolidScale Scan diagnostic includes a quick check.

What if my current AI provider does not have audit tools?

Ask the question explicitly, and ask what is planned for AI Act August 2026 compliance. If the answer is vague or absent, anticipate the change of provider before it becomes a regulatory emergency. The cost of anticipated migration is always lower than the cost of panic migration.

Does auditability slow down AI projects?

Not if it is integrated from the brief. It slows things down considerably if added after the fact on a system already in production. This is the main argument for treating it at the Scan stage rather than retrofitting.

Sources

• Anthropic Transformer Circuits, Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations, May 2026, transformer-circuits.pub/2026/nla
• European Regulation on Artificial Intelligence (AI Act), articles 13 and 50, artificialintelligenceact.eu
• Anthropic, Claude Opus 4.6 System Card, February 2026