AI Criminal Risk and Investigations
Criminal exposure in AI is not random. It concentrates in a handful of places — dual-use capability, data acquisition, and what a company does after a regulator tells it to act.
The risk is narrow, specific, and usually invisible until it is examined
Most AI founders are right that building a model is not a crime. They are wrong to conclude that the business carries no criminal exposure. The exposure does not come from the technology; it comes from three ordinary operational facts: the product can be used for a lawful purpose and a criminal one with equal ease, the training data was acquired in a way nobody documented, and a rectification notice arrived and was handled by someone in operations rather than by counsel.
Enforcement attention has shifted accordingly. Where earlier cases focused on the person operating the fraud, the questions now extend to the platform that supplied the capability, the API that was called at volume, and the service that stayed up after being told to change.
What we do
- Exposure mapping: function-by-function review of where dual-use capability, data acquisition and content handling create criminal risk
- Aiding-liability assessment: whether monitoring, rate limiting and response mechanisms are sufficient to negate knowledge
- Data acquisition review: lawful basis for collection, whether technical protection measures were circumvented, personal information content
- Compliance system build: monitoring and anomaly handling, content governance, access control, incident protocol
- Regulatory response: inquiry handling, investigation cooperation, rectification design and execution
- Remediation defence: documenting measures taken so that the record supports a due-diligence position
What is different for an AI company
- Data-related exposure is the most common. Scraping personal information, buying annotation data without checking provenance, or staff retaining customer data can each reach the personal-information offence; an API abused by fraud rings without remediation can engage liability for assisting cybercrime.
- Content-related exposure depends on product form. Duties around generating, storing and disseminating unlawful content vary by service type and knowledge; review and takedown processes must be producible on inspection.
- Training data can also touch copyright. Mass reproduction of protected works for training rarely meets the criminal threshold, but civil and administrative exposure usually arrives first.
- Corporate and individual liability run in parallel. The company being liable does not exonerate individuals; the legal representative and directly responsible managers can be pursued at the same time.
- Investigation response has a window. How the first interview, data production and seizure are handled shapes the eventual characterisation; designate a contact and prepare a document list in advance.
Typical situations
Situation 1: The API was used at scale for fraud, and the platform did not notice
A customer account calls the API in volume to generate material used in a fraud scheme. There is no monitoring rule, no anomaly alert, and no response procedure. The question at investigation is not whether the company participated, but whether it knew — or should have known.
Our approach: build the mechanisms that answer the knowledge question before it is asked — usage monitoring with defined thresholds, a prohibited-use policy that is enforced rather than published, a documented response when an alert fires. The value of these measures is evidentiary as much as operational.
Situation 2: Scraped data crossed a line nobody drew
Training data is collected by automated collection from sources that require login or that deploy technical protection measures. Whether the collection is lawful turns on details that were never recorded: what was accessed, under what terms, whether protection was circumvented, whether personal information was included.
Our approach: review the acquisition chain source by source, separate public-domain collection from access that required circumvention, and where the distinction is unfavourable, change the data source rather than defend the method. Provenance records built before collection are worth more than explanations built after.
Situation 3: A rectification notice was received and quietly shelved
A regulator orders corrective measures. The notice is forwarded internally, work is deprioritised, and nothing is filed back. Failure to perform a network security management obligation is constituted precisely by refusing to rectify after being ordered to — the step from administrative violation to criminal exposure is taken by inaction, not by the original defect.
Our approach: treat every notice as a dated obligation with an owner, a work plan and a written response. Rectification should be documented in a form that can be produced later, because the record of compliance is the defence.
Process and deliverables
| Stage | Work | Deliverables |
|---|---|---|
| 1 | Function-level exposure mapping, data chain review | Risk map + red-line register |
| 2 | Aiding-liability and monitoring assessment | Gap analysis + control design |
| 3 | Compliance system and incident protocol | Policy set + response playbook |
| 4 | Inquiry or investigation support | Response plan + submission documents |
| Ongoing | Enforcement tracking, system updates | Update notes + training materials |
AI-specific deliverables
- Model and data risk map: the criminal exposures attaching to training data, model weights and generated content, mapped to the conduct elements investigators examine.
- Inquiry-response protocol: who speaks, what is produced and in what sequence when investigators arrive.
Frequently asked
How is “knowledge” established in aiding cases? Judicial interpretation sets out indicia rather than a single test — continuing after being notified by a regulator, failing to act on a report, unusual pricing or settlement patterns, providing tools designed for the offence, and taking steps to evade supervision. The practical implication is that a company’s monitoring and response record is the evidence that decides the question.
Is a compliance programme a defence? Not a complete defence, and it should not be presented as one. It operates in two ways: it reduces the likelihood that the conduct occurs, and it supports the argument that the company did not knowingly lend assistance. Where a case proceeds, documented remediation is also relevant to how it is resolved.
We only provide technical capability — does that insulate us? Not by itself. The provision of technical support with knowledge of criminal use is the core of the aiding offence. What matters is whether the provider had knowledge and whether it acted. Neutral capability plus active monitoring is a materially different position from capability plus indifference.
When should counsel be involved in a regulatory inquiry? At the first written inquiry, not at the first interview. Early responses set the factual frame that later stages are argued within, and documents produced early are difficult to recalibrate.
When does a data problem become a criminal one? When the data was obtained by intrusion, bought from an unauthorised source, or used to evade identity verification. The threshold is how the data was obtained, not how much of it there is.
Does providing model capability attract liability? The exposure turns on knowledge and control. A general-purpose API with usage policies and monitoring is treated differently from a product designed for the conduct; keep records of what you knew and what you did.
Legal basis
- Criminal Law Art. 287-2: aiding information network criminal activities
- Criminal Law Art. 253-1: infringement of citizens’ personal information
- Criminal Law Arts. 285, 286: illegal access to and acquisition of computer information system data
- Criminal Law Art. 286-1: refusal to perform information network security management obligations, constituted by failure to rectify after being ordered
- Judicial Interpretation on Handling Criminal Cases of Illegal Use of Information Networks and Aiding Information Network Criminal Activities (2019): indicia of knowledge
- Cybersecurity Law and Data Security Law: the underlying management obligations whose breach feeds the offences above