{"id":11873,"date":"2026-09-09T10:28:02","date_gmt":"2026-09-09T08:28:02","guid":{"rendered":"https:\/\/inspace.io\/blog\/ai-hallucinates-the-mainstream-tmp-en-gb"},"modified":"2026-09-11T09:51:16","modified_gmt":"2026-09-11T07:51:16","slug":"ai-hallucinates-the-mainstream","status":"publish","type":"post","link":"https:\/\/inspace.io\/en-gb\/blog\/ai-hallucinates-the-mainstream","title":{"rendered":"AI Hallucinates the Mainstream"},"content":{"rendered":"

I spent a good while recently arguing with Claude about canonicals.<\/p>\n

The site had slot URLs, the kind of parameter-driven arrangement that does not sit neatly inside the textbook. I wanted a view on how to handle them. Claude had one. It had looked at the site, formed an opinion, and delivered it with the calm confidence of something that has read every SEO article ever published. Which is more or less what had happened.<\/p>\n

The answer was the mainstream answer. Canonical to the clean version, consolidate the signals, move along. Perfectly defensible in general. Wrong for this site.<\/p>\n

So I pushed. Not on the facts, on the judgment. I made it defend the recommendation against the specifics of the setup, and asked why it had never raised the option it had quietly ruled out. Two things came out of that. I got the right output. And it stopped treating its own first answer as settled. That second part is the interesting one.<\/p>\n

The advice going round this week<\/h2>\n

There has been a run of pieces making the same argument, and it is a reasonable one. Search Engine Land ran a piece arguing that AI works best when it removes work, not judgment, that it becomes an easy button precisely when you scope it to the tedious and repetitive, and turns risky the moment you let it decide. A week earlier the same outlet put it more bluntly: use Claude for SEO, do not let Claude do SEO. Lily Ray, arguing that shortcuts rarely win in SEO or AI search, arrives somewhere similar from a different door: the shortcuts have never won, and durable visibility comes from actual expertise.<\/p>\n

I agree with all of it. Not grudgingly, not as a floor to be improved on. Delegate the labour and keep the accountability is the right instinct, and it is a long way better than the alternative, which is a junior pasting an AI recommendation into a ticket because it sounded authoritative.<\/p>\n

I would just add the condition nobody states out loud.<\/p>\n

The advice works because it splits the job into a tedious half and a judgment half. Fine. But the tedious half is where the bad data lives. Redirect tables, product feeds, hreflang matrices, parameter handling, exports that were already wrong before anybody automated anything. Nobody looks closely at those, which is exactly why they got labelled tedious in the first place. Automate a mechanical task on top of an input nobody has audited and you have not removed risk. You have removed the person who would have noticed.<\/p>\n

So removing work is safe, on one condition: that the data underneath the work is sound. That condition is doing an enormous amount of load-bearing for something nobody checks.<\/p>\n

The leash<\/h2>\n

So the boundary is right. It just needs a hand on it. I run this as a leash, and I am not apologetic about the comparison.<\/p>\n

Four things have to be true before I let automation take anything meaningful. It gets instructions, so the job is specific rather than gestured at. It gets an Umfeld, a frame of reference to reason inside, because a model without a conceptual frame will default to the consensus every time. I need visibility into the grounds of its decisions, not just the output, because the output alone tells me nothing about whether it reasoned or recited. And then it gets room to move, but on a leash I can pull.<\/p>\n

The alternative is walking the dog with no lead at all and hoping it comes back when called. Sometimes it does. That is not a system, it is a mood.<\/p>\n

\"The
Room to move. Still attached.<\/figcaption><\/figure>\n

The distinction matters because it explains the canonicals episode. Claude did not invent a fact. It reproduced the mainstream position with total confidence and would not colour outside the lines. That is a hallucination too, just a socially acceptable one: a hallucination of consensus. Duane Forrester has described the brand-facing version of the same failure, where a model with thin information about your company confidently describes a competitor instead. Same mechanism, pointed at a different target. Thin input does not produce hesitation. It produces the average answer, delivered like a verdict.<\/p>\n

And an average answer is the one thing an experienced practitioner is never being paid for.<\/p>\n

The judgment you asked for, and the judgment that crept in<\/h2>\n

Which brings me to the line I would add to the week’s advice rather than argue with. I am not against giving AI judgment. I am against giving it judgment by accident.<\/p>\n

If you want judgment, ask for it deliberately and accept that you have bought a reliability problem, then manage it. If you do not want judgment, do not hand over a vague task and hope none creeps in. Give it a specific job. The failure I see most often is people wanting the second and asking for the first.<\/p>\n

\n\n\n\n\n\n\n\n\n\n
<\/th>\nYou want judgment<\/th>\nYou do not want judgment<\/th>\n<\/tr>\n<\/thead>\n
What you are actually asking for<\/strong><\/td>\nA recommendation between options<\/td>\nExecution of a defined job<\/td>\n<\/tr>\n
What you must supply<\/strong><\/td>\nContext, constraints, the Umfeld, the specifics that make your case unusual<\/td>\nAn unambiguous spec and a verified input<\/td>\n<\/tr>\n
What you have to accept<\/strong><\/td>\nIt is a reliability question, not a correctness guarantee<\/td>\nIt will do exactly what you asked, including when what you asked was wrong<\/td>\n<\/tr>\n
How you find out it went wrong<\/strong><\/td>\nYou interrogate the reasoning, not the answer<\/td>\nYou check the input data before you check the output<\/td>\n<\/tr>\n
The failure mode<\/strong><\/td>\nConfident consensus dressed as analysis<\/td>\nScaled, tidy, consistent nonsense<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Read the right-hand column twice, because it is the one the advice does not reach. You can keep every judgment call human, exactly as recommended, and still ship a redirect map built on a stale export. At volume. On schedule. Beautifully executed.<\/p>\n

Ask what it actually had to work with<\/h2>\n

There is a detail from this week that sharpens the point nicely. Data on how Claude and Claude Code use the web found that consumer Claude ran a web search in 93% of responses to a set of identical prompts, while Claude Code did so in 13%. Same underlying model. Wildly different grounding, decided by which window you happened to be typing in. The two also rarely name the same brands, which is a separate headache for another day.<\/p>\n

Sit with that for a second. Unless you knew that figure, you had no way of telling whether the answer in front of you came from something fetched five seconds ago or something absorbed during training and frozen since.<\/p>\n

This is the same gap I wrote about in Your LLM Doesn’t Know You, That’s Why Its Answers Are Average. An engine that lacks your context does not tell you it lacks your context. It gives you the averaged answer in a confident voice, and the confidence is not correlated with the grounding.<\/p>\n

What I would actually tell you<\/h2>\n

Stop asking whether AI is allowed to have an opinion. Start asking three duller questions, in order.<\/p>\n

What did I actually ask for, a judgment or a job? Did it have enough first-hand information about this specific situation to answer, or was it always going to hand me the consensus? And am I checking the reasoning, or just reading the output and moving on because it sounded right?<\/p>\n

\"The
Three duller questions, in order.<\/figcaption><\/figure>\n

None of that is a limitation on the tool. It is the part of the work that is still yours, and it is the part that eighteen years buys you. I got the right answer on those canonicals because I knew enough to distrust a textbook answer. The tool did not fail. It did what a confident, well-read, context-free advisor always does.<\/p>\n

Do not take my read on this as fixed, incidentally, on software that changes underneath you every fortnight.<\/p>\n

But the leash is not a lack of trust in the AI. It is trust in your own hand.<\/p>\n","protected":false},"excerpt":{"rendered":"

Let AI remove work, not judgment. Good rule, and I agree with it. It is also conditional, because the tedious half is exactly where unaudited data hides.<\/p>\n","protected":false},"author":7,"featured_media":11863,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_genesis_title":"","_genesis_description":"","footnotes":""},"categories":[449],"tags":[],"class_list":["post-11873","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"meta_all":{"rank_math_internal_links_processed":"1","rank_math_title":"AI Hallucinates the Mainstream, and Sounds Certain Doing It","rank_math_description":"Let AI remove work, not judgment. Good rule, and I agree with it. It is also conditional, because the tedious half is exactly where unaudited data hides.","rank_math_focus_keyword":"hallucination of consensus","rank_math_primary_category":"449","time_to_read":"7","meta_title":"AI Hallucinates the Mainstream, and Sounds Certain Doing It","meta_description":"Let AI remove work, not judgment. Good rule, and I agree with it. It is also conditional, because the tedious half is exactly where unaudited data hides."},"meta_all_flat":{"rank_math_internal_links_processed":"1","rank_math_title":"AI Hallucinates the Mainstream, and Sounds Certain Doing It","rank_math_description":"Let AI remove work, not judgment. Good rule, and I agree with it. It is also conditional, because the tedious half is exactly where unaudited data hides.","rank_math_focus_keyword":"hallucination of consensus","rank_math_primary_category":"449","time_to_read":"7","meta_title":"AI Hallucinates the Mainstream, and Sounds Certain Doing It","meta_description":"Let AI remove work, not judgment. Good rule, and I agree with it. It is also conditional, because the tedious half is exactly where unaudited data hides."},"acf":[],"_links":{"self":[{"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/posts\/11873","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/comments?post=11873"}],"version-history":[{"count":2,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/posts\/11873\/revisions"}],"predecessor-version":[{"id":11895,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/posts\/11873\/revisions\/11895"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/media\/11863"}],"wp:attachment":[{"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/media?parent=11873"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/categories?post=11873"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/inspace.io\/en-gb\/wp-json\/wp\/v2\/tags?post=11873"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}