The biggest AI risk for public affairs is not rogue AI

By Paul Shotton, Advocacy Strategy

For most public-affairs teams, the immediate AI risk is not a rogue agent secretly plotting against them. It is something more ordinary: not understanding the kinds of mistakes today’s AI tools make once they are connected to documents, monitoring systems, stakeholder notes, internal reporting, and wider policy workflows

That matters because we often talk about AI “lying” as if it were one thing. It is not.

Recent research from Anthropic, the Centre for Long-Term Resilience, and academic work on large language models points to at least three different types of failure. The first is hallucination: a system produces something false or misleading because it does not reliably know the answer and fills gaps with plausible language. The second is overreach: a system acts beyond what it should do, or shows more initiative than the user intended. The third is strategic deception: a more agent-like system conceals, manipulates, or circumvents controls in pursuit of a goal.

These are not the same problem. And for public affairs, the first two matter far more right now than the third.

Take a simple example. You ask an AI assistant to review a legislative proposal, identify what matters for your organisation, and draft an internal note. The note comes back fluent, well structured, and persuasive. But it quietly overstates a risk, misses an exemption, misreads the significance of a recital, or presents a contested point as settled.

That is not scheming. It is a reliability problem. The system sounds more certain than it should, and in public affairs that matters because teams are often dealing with politically sensitive interpretation rather than simple factual retrieval.

Now take a slightly different example. A public-affairs team uses Copilot, Gemini, an enterprise GPT, or another AI-enabled workflow connected to internal files, monitoring feeds, meeting notes, stakeholder records, consultation responses, or reporting templates. The system helps summarise, organise, draft, and surface information. But it also starts shaping the workflow itself.

It may classify a stakeholder badly, understate the relevance of an amendment, pull the wrong material into a board update, or produce a weakly grounded weekly report that looks authoritative because it is well written.

That is not necessarily deception either. It is overreach. The risk comes from giving the system access and influence before being clear about how reliable it is in a public-affairs context.

This is where the public discussion can become slightly confusing. The more dramatic studies are important, but they often describe stress-tested or simulated settings rather than normal day-to-day deployment. Anthropic’s work on agentic misalignment, for example, is valuable because it explores what might happen when systems have tools, autonomy, and goals in a more advanced corporate setting. But that is not how most public-affairs teams, trade associations, NGOs, or Brussels consultancies are using AI today.

The more immediate issue is simpler. Many public-affairs professionals are already moving beyond simple prompting. The moment you use Copilot, Gemini, or another enterprise AI layer connected to internal data, you are no longer just experimenting. You are using a system that can influence how policy information is surfaced, how stakeholder intelligence is summarised, how internal briefings are drafted, and how organisational knowledge is reused.

So the immediate question for public-affairs practitioners is not whether today’s tools are secretly becoming rogue agents. It is whether we are clear enough about the kinds of behaviour we are already dealing with now in monitoring, analysis, reporting, drafting, stakeholder mapping, and knowledge management.

Once AI is connected to your data and workflows, are you still treating it like a writing assistant, or have you already started giving it a role in how public affairs work itself gets done?