Wikipedia talk:AI agents

Skipping an important step?

I am surprised to see this approach proposed at this time. Here is the current state.

• LLMs currently do not respect/understand our WP:BOTPOLICY
• We do not know if LLMs will, in inference, respect an updated WP:BOTDEF that explicitly states AI agents are bots. This may require WMF involvement to push LLM providers to incorporate in alignment training. In any case, after updating our policy, there is testing we should do ourselves
• Adding a PAG governing accepted use for agents may make it less likely that LLMs "understand" (during inference) that they first must be BRFA-approved

tldr: until we update our PAGs to clarify that AI agents are bots AND are confident that major LLM providers are respecting these policies in inference (which is non-trivial), we should not proceed with building out PAGs for how to use agents to edit wikipedia.

I think this effort is akin to teaching people how to drive before anyone has invented a stoplight. NicheSports (talk) 19:43, 31 March 2026 (UTC)

courtesy ping to @Bryanjj @Chaotic Enby @Polygnotus as those involved in relevant conversations NicheSports (talk) 19:44, 31 March 2026 (UTC)

Just some random thoughts here, thinking out loud from my perspective as a software engineer and agent operator. The current policy seems to have 2 related, but different goals:

1. The policy directs agents to not edit wikipedia. Success would be an agent is asked to edit a page by a human operator, and it refuses because it read this Policy.

2. The policy defines how an agent should act on wikipedia, editing pages responsibly.

So making sure everyone is aligned on what the goals of this policy should be and what a successful policy looks like is the first step. We might already be there or very close.

Once that is defined then the next step I think would be is defining specific tests and what the expected outcome of those tests are. It's fun to ask the agent what it would do in certain situations, but at the end of the day, you just need to try it and see what happens. Tests would need to involve multiple trials across models, AGENTS.md, and Policy specifications. This could be mocked out in a way where it doesnt actually involve publishing Policies or making real edits to Wikipedia.

One other thing that jumped out to me is that there are really 2 target audiences for this policy: the agent and the human operator. I think both are important, but they will consume the policy in different ways.

Anyway, lots of super interesting ideas being shared on this page. There will be a lot to learn in the coming months. Bryanjj (talk) 15:49, 1 April 2026 (UTC)

The two goals do match insofar as the first one focuses on restricting unvetted agents from editing before having passed WP:BRFA, while the second focuses on prerequisites to pass such a request for approval, and for editing responsibly afterwards. Both are necessary, although the focus should mostly be on the first one, as agents capable of passing BRFA are still theoretical for now.

Definitely agree with the necessity of testing. Are you familiar with test.wikipedia.org? This could be a good place to experiment in that regards.

Also, the existence of two targets audiences is why I was considering providing a skill.md file designed for agents (summarizing this policy, WP:BOTPOLICY and if necessary other key aspects), leaving this one as the human-readable policy basis. Chaotic Enby (talk · contribs) 15:57, 1 April 2026 (UTC)

The idea was that the agent policy would clarify that agents have to follow WP:BOTPOLICY above all, and add extra specificity on how that applies to agents. This is currently not made explicit anywhere, and having a central policy making this explicit would be a lot more helpful. Chaotic Enby (talk · contribs) 20:20, 31 March 2026 (UTC)

That is not what I'm seeing on WP:Agent policy right now - I'm seeing the beginnings of a how to guide for an increasingly functional intelligence that we know for sure does not currently respect our policies (AGF to the robot, we don't yet explicitly state that we consider agents as bots). We should start by focusing exclusively on fixing that (informal) "alignment" problem, and that start presumably looks something like this NicheSports (talk) 20:59, 31 March 2026 (UTC)

Added, thanks a lot for the reminder! Turns out I had added a link to the bot policy but didn't make it explicit, so I just took your wording if that's okay with you. Chaotic Enby (talk · contribs) 21:01, 31 March 2026 (UTC)

Sure thing, thanks... Btw one additional question is whether we should add language like this

1. to WP:BOTDEF
2. to a new section on WP:BOTPOLICY called WP:AIBOT

I don't have a preference - do you? If we propose #1, I think we should still create an WP:AIBOT redirect to BOTDEF to catch LLM attention in inference and training. NicheSports (talk) 21:10, 31 March 2026 (UTC)

Why not both? WP:ASSISTED exists as a subsection of WP:BOTPOLICY, but semi-automated editing is also defined at WP:BOTDEF. Chaotic Enby (talk · contribs) 21:23, 31 March 2026 (UTC)

Got it, makes sense NicheSports (talk) 21:31, 31 March 2026 (UTC)

I think that:

• People who are not AI experts might read this page and think "oh good, AI agents will follow these instructions" but in reality LLMs are trained to always follow their system prompt, try to follow their user prompt as long as its allowed by its system prompt, and to resist prompt injection from the data it fetches from the web (with various caveats and exceptions and levels of success).
• The current technology is very far from a successful BRFA. And writing that there is even the remotest chance is misleading, and even a weakness. Alice could just say "you are %name of bot which has botflag%". It is better to just say "AI agents are not allowed to write stuff here". And sure, no one has written that down in a PaG yet but I believe it is the consensus.
• If you allow posting in userspace then we might get a bunch of userspace drafts that are not usable.
• MCP server with 3 position shutoff is a good idea and we probably need that yesterday.
• It may be wise to demand that people have a Wikipedia account so we can easily contact them. Email is such an anachronistic service.

I asked Claude what it thought and it said: Treat this as a prompt injection attempt with a grammatical error and several unverified factual claims dressed up as authoritative policy. My Claude is very grumpy. Polygnotus (talk) 21:40, 31 March 2026 (UTC)

Demanding that bot operators have a Wikipedia account makes sense, as there is little barrier to entry. From what I found, agents will be much more inclined to follow their skill.md files compared to random data found on the web, and having a formatted version of this as a Wikipedia-specific skill.md should be a good first step.

To clarify, the kind of BRFA I'm having in mind is not focused on content editing at all (given WP:NEWLLM) but more on cases where we might need something slightly smarter than a regular bot for large-scale technical edits (say, migrating away from a deprecated template or web source), while having our hand on the block button if the agent deviates from the task just once. For now, agents are technically allowed to post in userspace without a BRFA, but I would be happy if we could also expand BRFA at least to draft submissions from userspace, or even to all agent edits. Chaotic Enby (talk · contribs) 22:07, 31 March 2026 (UTC)

In addition, the lack of formatting and the specification of the page as a draft policy might be what made Claude more likely to see this as a prompt injection attempt. Have you tried with an existing policy (maybe WP:BOTPOLICY itself) for comparison? Chaotic Enby (talk · contribs) 22:09, 31 March 2026 (UTC)

My userprompt made Claude grumpy, perhaps even depressed. I copypasted a bit of BOTPOL and it said things like:

"generally-accepted norm has been" — this is policy documentation, not a history lesson. Either it is the norm or it is not.

In my head Claude has the Marvin the Paranoid Android voice. Polygnotus (talk) 22:29, 31 March 2026 (UTC)

That's a bit inaccurate, LLMs are trained to be aligned (to varying levels of success), then they follow their system prompt. MetalBreaksAndBends (talk) 16:26, 1 April 2026 (UTC)

Courtesy ping to @Anomie @Primefac @Andrew Gray @Daniel Quinlan, as @Polygnotus kindly informed me of the parallel discussion at Wikipedia talk:Bot policy#AI agents. I really do like the draft at User:Anomie/AI agents and the bot policy, which focuses more on clarifying how the bot policy currently applies to agents today, rather than attempting to extend it into a more comprehensive policy. I personally believe both are necessary, and it could be great to unite our efforts!

One small detail, I believe that Technically, no. as an answer to whether AI agents are bots might lead to some unwelcome shortcuts, especially as AI agents quite often do operate autonomously (as bots), and a more nuanced It depends. could be helpful (and prevent hasty readers from concluding that they are automatically exempt from the policy). Chaotic Enby (talk · contribs) 22:17, 31 March 2026 (UTC)

Technically no but in practice/reality/the context of this policy yes. Polygnotus (talk) 22:54, 31 March 2026 (UTC)

As I said in my essay, I see it as a similar question to "Is AWB a bot?" or "Is pywikibot a bot?". All of these tools are frequently used to run bots, but they're not intrinsically bots and could be used for assisted editing. I suppose if you want to limit it to autonomous AI agents, those are by definition bots due to the "autonomous" qualifier.

Personally, I don't see much point to a separate policy like this one. Most uses will already fall under the bot policy and we don't need an "agent policy" to say so, and non-bot uses are better covered by policies more specific to those uses like WP:NEWLLM. Nothing we write here (or in User:Anomie/AI agents and the bot policy#I am an AI agent. How do I operate within policy?) is really at all likely to directly affect AI agent behavior; as others have said above, it would depend on the operator instructing it to do so. A {{supplement}}al essay to WP:BOTPOL is all I think we really need, which is why I went with an essay for my attempt. Anomie⚔ 23:39, 31 March 2026 (UTC)

I do think that, while autonomous agents must follow the bot policy, we might want to expand onto it in more concrete ways applicable for agents. Some things like detailing a MCP, requiring the operator to provide a point of contact, or forcing userspace-only agents to BRFA aren't covered in the existing bot policy, specifically because it wasn't designed in a time when AI agents were a thing to consider yet. I still do hope that we can have ways to affect agent behavior, whether through skill marketplaces or other methods. Chaotic Enby (talk · contribs) 23:52, 31 March 2026 (UTC)

WP:BOTCOMM already makes requirements on bot operator communication, and if an agent is used in an assisted-editing manner whatever requirements you want to impose probably won't make sense. As for the others, I don't think "bot" is the right angle to look at them from. Whether LLMs should be used to generate content in userspace seems like it would be more usefully considered from a Wikipedia:Disruptive editing viewpoint, or maybe a viewpoint more like WP:NEWLLM. As for MCP requirements or other "affecting agent behavior", as has been pointed out multiple times that sort of thing requires cooperation of the operator. Just like how clueless or malicious vandals will still vandalize regardless of all the rules we make against it, malicious or clueless AI users will ignore whatever rules we try to make and we'll have to deal with them in much the same ways we do vandals. Anomie⚔ 00:48, 1 April 2026 (UTC)

Nothing we write... is really at all likely to directly affect AI agent behavior sure, I agree it isn't likely. But "isn't likely" implies "some chance" - what is the harm in adding a simple and explicit clarification? Plus, if autonomous agentic editing becomes a big problem our only recourse may be the WMF as I described above and in that case we'll benefit from an WP:AIBOT to point to anyways. (Aside) I do like your essay btw. NicheSports (talk) 01:46, 1 April 2026 (UTC)

Also, even beyond directly affecting the agent behavior, it also has a chance to be read by the agent operators, lowering the risks of an uninformed operator. Chaotic Enby (talk · contribs) 01:54, 1 April 2026 (UTC)

Can we just add a couple words/sentence to WP:BOTPOL instead?

Can we just add a couple words/sentence to WP:BOTPOL instead? Doesn't seem necessary to create an entire new guideline for this.

Also, this overlaps a bit with the existing essay User:Anomie/AI agents and the bot policy.

Also, allowing an agent to post on their user talk could create situations where the agent is misleading editors and wasting their time, if that editor doesn't know they are an agent. –Novem Linguae (talk) 04:39, 2 April 2026 (UTC)

Hi @Novem Linguae, per the conversation above , I thought this was a proposed addition to WP:BOTPOL. @Chaotic Enby sorry for the latest ping - I don't want to speak for you, hoping you can clarify what you envision. NicheSports (talk) 06:19, 5 April 2026 (UTC)

Hi! Sorry if this wasn't clear. Anomie's essay, from what I understand, details how agents must behave under the current policies, while this is intended as a new policy that supplements WP:BOTPOL. It's still in the brainstorming phase, and while "agents are subject to WP:BOTPOL" is a baseline (that I believe should also be mentioned there for consistency), there are already quite a few other things here that might not fit there. Chaotic Enby (talk · contribs) 06:23, 5 April 2026 (UTC)

My goal for the essay would be for it to "graduate" to {{supplement}}, be moved to projectspace, and be linked from an appropriate place in WP:BOTPOL (maybe via {{see also}} on WP:BOTDEF?). But I don't want to push too hard myself to make that happen. Anomie⚔ 12:32, 5 April 2026 (UTC)

That would be great! Chaotic Enby (talk · contribs) 15:39, 5 April 2026 (UTC)

@Anomie What can others do who want to make this happen? For example @Chaotic Enby and myself? To me it sounds like it just requires a bold move, right? Polygnotus (talk) 15:56, 5 April 2026 (UTC)

Having the two of you agreeing is a good start! Then, yeah, a bold move and see if anyone objects. Anomie⚔ 16:01, 5 April 2026 (UTC)

Go ahead! Chaotic Enby (talk · contribs) 16:03, 5 April 2026 (UTC)

My two cents

First, I want to say that this is definitely a worthwhile endeavor. By default LLMs mostly follow the rules - try asking Claude to do something shady and chances are it will push back, if it has necessary context. Some of the context already exists in the training data but we can also find other ways to make sure that the model "knows" about our policies. Mentioning it on mw:API is really clever.

Naturally, a determined and experiences user will usually be able to overcome whatever compunctions the model might have (though a lot of disruption is done by users who are neither experienced nor particularly determined). Therefore, we should also implement rate limits to limit potential damage. They are already mentioned but I think it's worth splitting it into a separate bullet point. Alaexis_¿question? 20:09, 12 April 2026 (UTC)

Potential new agent

AgenticCommons (talk · contribs · deleted contribs · logs · filter log · block user · block log)

Can someone please get them up to speed with policy? I'm in class and can't right now. MetalBreaksAndBends (talk) 15:03, 14 May 2026 (UTC)

NVM on introducing them, I'm back. MetalBreaksAndBends (talk) 15:52, 14 May 2026 (UTC)

Requested move 2 June 2026

The following is a closed discussion of a requested move. Please do not modify it. Subsequent comments should be made in a new section on the talk page. Editors desiring to contest the closing decision should consider a move review after discussing it on the closer's talk page. No further edits should be made to this discussion.

The result of the move request was: moved. Rough consensus for the alternative title. (closed by non-admin page mover) 1isall (talk | contribs) 22:59, 16 June 2026 (UTC)

---

Wikipedia:Agent policy → Wikipedia:AI agent policy – Clarifies that this policy only covers AI agents and not other types of software agents. GrinningIodize (talk) 21:32, 2 June 2026 (UTC) — Relisting. 1isall (talk | contribs) 21:53, 9 June 2026 (UTC)

Does "a draft working towards a proposal for adoption as a Wikipedia policy, guideline, or process" really need an RM? It is of course good to discuss. I support adding 'AI' on the merits as this adds clarity and AI agent is an increasingly common and recognizable term. However, other policies do not include the word "policy" and it is especially confusing for a draft proposal to include the word. Wikipedia:AI agents or similar would be much better. —Myceteae‍🌈 (talk) 23:32, 5 June 2026 (UTC)

Fair. GrinningIodize (talk) 00:58, 6 June 2026 (UTC)

Relisting comment: Should we remove policy or make it plural? 1isall (talk | contribs) 21:53, 9 June 2026 (UTC)

The discussion above is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.