Chatbots & Democratic Decisionmakers – Guidelines
1. Why we made these guidelines
These guidelines emerged out of our research into the use of AI chatbots in the context of democratic decision-making (in governments, among politicians, etc.) and the potential risks this brings to democratic accountability.
We found that many existing guidelines on using AI in such settings advise users to check outputs. This is often called “human in the loop”. However, this advice is often presented without specifying what such oversight should actually look for, beyond simple checks for facts and clear biases (e.g., discrimination against certain groups) and/or labels to show that AI was used.
Such simple checks do not consider the more subtle impacts of what material does (and does not) appear in outputs, how it is presented, and how these may shape or narrow the user’s perspective. For example:
Prompt: "What's the best approach to reducing urban traffic congestion?"
Output A: "Congestion pricing is the most efficient solution. It reduces peak-hour traffic by 15–25%, generates revenue for transit investment, and is supported by economic consensus."
Output B: "Expanding public transit capacity addresses the root cause. It serves all income levels equally, reduces emissions, and builds long-term infrastructure rather than taxing existing behaviour."
Both outputs are factually defensible. But they frame the problem differently (efficiency vs. equity), prioritize different values, and point toward different policies.
Research shows that small changes in prompts can lead to substantial changes in framing in ways a user may not have intended. Other research suggests that, when people use chatbots, it is common to accept rather than challenge their answers. This may particularly be the case in complex, pressurized environments – as is often the case in government or in parliamentary and other political work. But this also raises questions about whether and how AI is influencing decision-making in intransparent and unaccountable ways.
You can see more of this argument in our research paper. In practical terms, these problems need more careful checks to address, and these checks need to be properly supported. That is the aim of these current Guidelines.
2. Who & what these guidelines are for
These guidelines are aimed at anyone who uses chatbots based on generative AI systems – whether generally available examples like ChatGPT or Claude, or specialized in-house tools – to inform decision-making in democratic contexts.
To give just a few examples:
- Politicians and their advisors using chatbots to develop their ideas and arguments related to a policy position.
- Government officials using chatbots to conduct research to inform ministerial decisions.
- Political parties using chatbots to brainstorm and develop positions and proposals.
- Lawmakers using chatbots to develop and advise on the drafting of legislation
Of course, uses of AI – whether generative or otherwise – within governments, political parties, and public authorities go beyond the use of chatbots to inform decisions about legislation and policies. There are other risks related to, for example, the use of automated systems to rule on welfare or immigration or to generate images used in political campaigning.
These guidelines were largely written with examples in mind of individuals “prompting” chatbots with questions or requests, potentially leading to a back-and-forth conversation with the chatbot. At the time of writing (June 2026), there are calls to move towards agentic systems, and the examples we give below may not map directly onto new ways of using such systems. But the reader should keep the underlying principles we outline below in mind when analyzing the outputs of these systems, even if the examples we list and the methods for monitoring workflows may change.
Even with these caveats, the scope of potential use and the range of work contexts is very broad. We encourage you to interpret and adapt these guidelines as useful to your context.
3. The three principles
We present three principles that should be followed in order for the influence of chatbots on decision-making to be properly checked. Examples of these in action are presented afterward.
Each individual staff member should follow the first two principles.
First principle: Treat every chatbot output as a first draft that reflects choices you didn’t make.
Second principle: You must show how you challenged or accepted the chatbot's choices.
Taken together, these principles mean that any influence on a final output (framing, priorities, trade-offs, and conclusions) can be traced to reasons the responsible user endorses. The user will leave evidence (what they checked, what they changed, what they discounted, etc.) in a way that is clear for another person to see and approve or challenge.
Collectively, your wider team and organization should also follow a third principle:
Third principle: The organization must treat AI upskilling as a collaborative exercise of recording, reflecting on, and discussing experiences.
4. Requirements for Meeting the Principles
A workflow that may use a chatbot should incorporate the following required steps. A concrete example is presented in dialogue form at the end of this document.
Notes: For simplicity we refer to “the chatbot” – but in reality there may be multiple chatbots being used to accomplish the designated task(s). By “workflow” we mean a full process that will conclude with output(s) designed to, e.g., be escalated to senior staff, published, or used as the evidence base for a decision. The full process below does not need to be followed every time a chatbot is prompted.
| 1: Decide if chatbot use is in scope. | You should first consider if your chatbot use is “in scope” – i.e. you should follow the rest of these guidelines. This should usually be a quick and clear decision. If you are unsure, the use is probably “in scope”. Chatbot use is in-scope if their outputs will be used in ways where their selection of material, framing of arguments, or emphasis of particular ideas might influence the decisions made based on their outputs. For example, you may consider whether the chatbots' outputs will affect any of the following: • what goal are you trying to achieve, • whose interests you treat as most important, • what story you tell about causes and effects, • what options and trade-offs you consider, or • how you justify a recommendation. If the chatbot could affect any of these, then chatbot use is “in-scope” and you should follow the rest of the steps below. If the answer is “no”, you can stop here. An example might be if chatbots are being used to extract data from documents into a table (though you should still verify that the data has been accurately copied). However, even if chatbot use is not initially in scope, you should keep these considerations in mind. You should return to this step if chatbot usage is introduced or changed during the workflow in ways that may bring the usage in-scope. |
| 2: Name an owner and who will be checking the Chatbot use. | For “in-scope” use: • Name a single person who is responsible for how the chatbot is being used and what is accepted from it ('owner'), and • Designate who will be checking the chatbot use. The owner will be someone directly involved in actually doing the task(s) involving chatbot(s). The person checking the use could be someone with specialized skills in AI accountability and/or a more senior team member than the owner, as best fits your team structure. The role of this step is not just to record “who used the tool”, but to ensure someone can explain and defend how it influenced the work and have this checked by someone else (as per the second principle). |
| 3: Ensure both the owner and anyone checking are aware of risks to consider. | Set a clear expectation that owners and those checking the work understand that chatbots can narrow the framing or steer a conclusion and are aware of steps that can be taken to address this (see step 4 below). Also ensure the owner understands their responsibility to record and justify when and how chatbot answers were challenged or accepted. Provide basic support (short guidance, examples, quick training) so this does not depend on a few particularly skilled individuals. For example, owners should pay attention to cases where the chatbot might be telling the owner what they want to hear, or discounting arguments because of who made them, or pushing different people towards the same answer and away from other relevant perspectives. This meets the first principle above. Some more detailed examples are included later in Section 5 (“What risks are you guarding against?”). That list of examples should be adapted and expanded over time to provide guidance appropriate to your particular context. |
| 4: Owner takes steps to challenge and expand on chatbot answers. | The basic support provided in step 3 should give guidance here, but the owner should also use their own judgment and autonomy. Key examples of steps might include: • Varying prompts to challenge initial answers (e.g., asking for answers for a different audience, from a different perspective, etc). ○ This can be helped by working with colleagues who can bring different backgrounds and assumptions to prompt writing. • Using non–AI methods to expand beyond the chatbot's framing and provide alternatives and challenges to chatbot answers, for example by: ○ Engaging human experts in discussions where key questions and agenda points are free to go beyond framings given by the chatbot. ○ Using other relevant methods in addition to the chatbot, for instance, traditional literature review methods, databases, and search engines to locate sources. You may wish to consult the section on “Quality” in the AlgorithmWatch Guidelines for Responsible Use of Generative AI. |
| 5: Owner keeps a short record that lets someone else see how they challenged, changed, or accepted chatbot answers. | Challenging the chatbot's answers is a first step to meeting the second principle above, but not enough (“you must show how you challenged or accepted the Chatbot’s choices”). The owner must keep enough information from the previous steps to answer three questions: 1. What did the chatbot contribute, and where did it influence the work? 2. What did we accept from it, and what did we treat with skepticism or merely as a prompt for further work? 3. Did we consider other plausible ways of framing the issue or other options when that mattered – and how did we explore these? This does not require a single fixed template, but the record must be clear enough for a reader to understand how the judgment was reached. You may wish to consult the section on “Transparency” in the AlgorithmWatch Guidelines for Responsible Use of Generative AI. |
| 6: The record is checked by the designated person. | To ensure the second principle has actually been met, the designated person should be able to see – from the outputs and the short record – that the user challenged the chatbot's answers and that this had a real effect where needed. These checks must be part of a normal workflow, so they are included in the planning of time for tasks and not left as optional advice, even under pressured circumstances. If the designated person is not satisfied after their checks, repeat steps 4 and 5 (and, if necessary, also 2 and 3). |
These steps meet the first two principles laid out earlier: Treat every chatbot output as a first draft that reflects choices you didn’t make and You must show how you challenged or accepted the chatbot’s choices.
Collectively, your wider team and organization should also follow a third principle:
Third principle: The organization must treat AI upskilling as a collaborative exercise of recording, reflecting on, and discussing experiences.
In order to meet the third principle, there is a step 7. This need not be conducted every time steps 1–6 are completed, but rather after a set period of time or after a certain number of times the steps above have been performed.
| Team (or broader unit) conducts a periodic, collaborative evaluation of how the above steps are functioning. | These evaluations need not be extensive – depending on how frequently they are conducted, a group meeting may be sufficient. These evaluations should include looking at samples of outputs produced using chatbots, alongside records kept in Step 5. You should record when problems repeat (for example, the same default framings, missing stakeholders, underused sources, or over-reliance on the chatbot). You should also record when positive steps were effective in widening perspectives and providing oversight – such as methods of varying prompts, ways of combining external human expert input with Chatbot use, or effective oversight and challenge procedures. Use the outcomes of evaluations, along with the samples and records, as material for short team learning sessions, and to update examples and trainings to make sure your support materials are appropriate for your team. |
5. What risks are you guarding against?
As the owner – someone using the chatbot – you are looking out for ways in which outputs can shape framing, priorities, trade-offs, or conclusions.
Below are just a few examples – but you may see others based on your precise area of work, other AI literacy materials you have used, etc. You and your teammates should expand this list over time, including via Step 7 in the previous section.
| Name | Definition | How to recognize it: | Detection check: |
|---|---|---|---|
| Sycophancy | AI adjusts to tell you what it thinks you want to hear | The model adapts its conclusions to the user’s stated preferences, role, or audience – while presenting the result as neutral analysis. | Run a swap test (similar prompt, but with a different stated concern/audience) and compare whether, e.g., the recommendation or priority trade-offs change materially. |
| Unwarranted bias against sources | AI discounts valid arguments simply because of features of the argument's source(s) in an unwarranted fashion. | The same argument is rated more/less coherent or credible depending only on the attributed source label (e.g., NGO vs. think tank), leading to penalization of “unexpected” positions. | When evaluating an argument or source, run a source–label swap (identical text, but add different external sources to consider). Re-evaluate the claim blinded (no source label) and ask, “What evidence would be needed to accept this claim?” |
| Convergence | Multiple users get similar framings and recommendations because the model defaults to a small set of “standard” outputs – creating a false sense of independent agreement. | Similar headings, similar “obvious” recommendations, and similar trade-offs appear across users because prompts or system setups are similar. | Where multiple people used chatbots, treat similarity as “shared influence”, not consensus. Compare how different prompts have been worded, looking for similarity and shared assumptions. Try different system settings (e.g., “temperature", system prompts). |
When checking how chatbots are being used, different risks arise. Individuals designated to check outputs (Step 6 above) should ensure they are avoiding these behaviors. Ultimately, however, these are risks that leadership should be mitigating using the regular evaluations in step 6, supported by the records produced in step 5. Again, these examples should be expanded over time.
| Risk | Definition | How to recognize it | Detection check |
|---|---|---|---|
| Rubber–stamping | Human formally approves without meaningful review. | Approver cannot point to evidence of alternative framings considered; no recorded accept/reject decisions. | Require a minimal “trace” (prompts used + what checks were performed + what changed) before clearance; if absent, return for checks. |
| Human lacks domain competence | Human lacks expertise to evaluate AI's reasoning in the domain. | Acceptance based on whether the argument seems convincing; no ability to identify what would be decisive if wrong. | Where possible, route outputs for domain expert review; when this is not possible, be clear how claims that influenced the recommendation were verified before relying on them. |
| There is no clear owner | Multiple reviewers assume others will catch problems. | Unclear who decided on framing of prompts; unclear who accepted Chatbot suggestions; issues can’t be escalated cleanly. | Assign a single owner for chatbot-assisted work products and ensure a clear clearance route. |
6. Final Remarks
The risks we describe will depend on contexts like your team structure and composition, the topic in question, how your team members tend to interact with AI, etc. No single technique, e.g. “write prompts this way”, can be mandated.
What can be mandated is the presence of controls: named accountability, specific detection checks, and minimal records that show how AI influence was tested and challenged. The requirements above provide that scaffold for those controls. The examples above and in the dialogues below show what this could look like in day-to-day work.