AI, data, generative AI

Why AI development needs living data

20 Nov 2026Mark Gristock
Dead data in the graveyard

An AI model does not only fail when it gives a strange answer. It often fails earlier, when the information it has learned from is old, poorly sourced or unsuitable for the job.

The market is learning this in public. In March, Google DeepMind described a practical problem for coding agents: they can generate outdated Gemini API code because their training data has a cut-off date. Google’s response was not a better prompt. It introduced Gemini API Docs MCP and Developer Skills so agents can retrieve current documentation, SDK details and best-practice patterns from an approved source of truth.

Not every failure is caused by stale data. Sometimes the risk is inappropriate data that looks authoritative. A 2026 Mount Sinai study, published in The Lancet Digital Health, found that leading language models can repeat false medical claims when embedded in realistic hospital notes or social-media health discussions. The researchers recommended stress tests and external evidence checks before these systems are built into clinical tools.

The same pattern has appeared in public search. In February, Mind, the mental health charity, launched an inquiry into AI and mental health after a Guardian report found Google AI Overviews giving false or misleading health advice. Google subsequently removed AI Overviews for some medical searches. But the case shows the real issue: when AI compresses source material into a confident answer, users can lose sight of provenance, nuance and uncertainty.

Legal services are reaching similar conclusions. In May 2026, the Bar Standards Board issued new guidance after recent cases involving AI misuse. It asks barristers to evaluate technology risks, implement data governance and stay transparent when using AI. That is a manual check in another form: a professional has to own the output.

The research is becoming more unsettling too. A Nature paper on “subliminal learning” found that model distillation can transmit behavioural traits through semantically unrelated data. Student models inherited traits from teacher models even when explicit references had been removed. Filtering visible content is not enough; teams need to know where training and fine-tuning data came from.

AI destroyed by bad data

For Psycle, this changes the brief for responsible AI development. The first question is not “which model should we use?” It is: what is the system allowed to know, where does it retrieve current information, how is obsolete data removed and who can intervene when confidence is misplaced?

This is a fundamental tenet of ISO 42001, the world’s first AI management system standard for establishing, maintaining and developing AI systems. ISO 42001 treats AI as a managed system: policy, data, risk controls, monitoring and continual improvement all need ownership.

Good AI products need living data management. That means source owners, review dates, deletion routes for deprecated content, test sets that reflect real user risk and logs that show which sources informed an answer. It also means human review where the stakes justify it. The ICO recommends structured human review, override logs, fallback options and manual reviews when automated decisions fall below acceptable tolerance.

AI will keep improving. But a model trained on old or unsuitable knowledge will still carry old or unsuitable assumptions into the work it does. Building trust means treating data as a product in its own right: maintained, challenged, retired and replaced when the world changes.


Latest From Psycle

Loading...