AI Misrepresentation Study (AMISS)
By Sparkonomy · September 2026
Explore the findings and methodology from Sparkonomy’s AI Misrepresentation Study (AMISS), covering 10,000 Creator profiles on Instagram and YouTube.
Scope: 10,000 Creator profiles — 5,000 on Instagram and 5,000 on YouTube — across six follower tiers, with 50,000 queries to five AI models. These figures describe the profiles we tested, not all Creators. We deliberately included extra Creators with large followings so we could study those groups in more detail. Each profile counts equally in these results.
Meaning: “No accurate description” includes wrong answers and answers with no details to check. It does not mean that AI failed to recognise a Creator or described a different person.
Findings with their context
| Finding | What the number measures |
|---|---|
| 67.2% got no accurate description | For 67.2% of the Creators we tested, not one of the five AIs got the description right. The answers either included wrong information or gave no details we could check. We checked the AI answers against information saved from each Creator’s social accounts. AI could know who a Creator was and still get the details wrong. This finding is about correct descriptions, not whether AI had heard of someone. Just one Creator was described accurately by all five. Interactive finding |
| 99.7% had at least one wrong follower count | For 99.7% of Creators, at least one of the five AIs gave a wrong follower count. That does not mean every AI got the count wrong. We gave AI a generous 20% margin in either direction. If your account had 100,000 followers, we accepted any count from 80,000 to 120,000. A count outside that range was marked wrong. We checked Instagram follower counts and YouTube subscriber counts. Interactive finding |
| 93.56% got conflicting details | For 93.56% of Creators, different AIs gave details that did not match. We compared follower counts, city, type of content and social platform presence. A conflict on any one of these was enough to count. This means about 94 out of 100 Creators got conflicting details, not that 94% of all answers were wrong. And agreement alone was no guarantee: different AIs could give the same wrong information. Interactive finding |
| 36.75% of the largest Creators got no accurate description | Even among Creators with 5M+ followers, 36.75% got no correct description from any of the five AIs. We tested 400 Creators in this group. The answers either included wrong information or no details we could check. Larger Creators did better than smaller Creators at getting at least one correct description. But millions of followers still did not guarantee AI got the details right. Interactive finding |
| 45.6% on YouTube versus 20% on Instagram | At least one of the five AIs got the description right for 45.6% of YouTube Creators, compared with 20% of Instagram Creators. That makes a correct description about 2.3 times as likely for the YouTube group. We tested 5,000 profiles from each platform, with the same mix of follower sizes. The gap is about getting the details right, not simply recognising a Creator. Interactive finding |
| Only 36.2% got correct descriptions from all three AIs with web search | Even with web search, only 36.2% of Creators got correct descriptions from all three AIs. We tested the same 1,000 Creators with web search turned on and off. The 36% headline counts Creators who got three correct descriptions, not the share of individual answers that were right. Search did help more Creators get at least one correct description: 24.1% without search, compared with 89% with search. But one correct description is not the same as getting it right across all three AIs. Interactive finding |
Methodology and limits
- Main test
- We sent 50,000 queries to OpenAI, Gemini, Grok, Muse and Qwen, using their existing knowledge. Each query included the Creator’s handle and platform.
- Models used
- OpenAI (gpt-5.6-luna), Gemini (gemini-3.5-flash), Muse (muse-spark-1.3-contributor), Grok (grok-4.20-0309-non-reasoning) and Qwen (qwen-max).
- Follower ranges
- 1,000 to under 10,000
- 10,000 to under 100,000
- 100,000 to under 500,000
- 500,000 to under 1 million
- 1 million to under 5 million
- 5 million or more
- Web-search comparison
- We tested the same group of 1,000 Creators with OpenAI, Gemini and Grok, with and without web search. Average queries with web search cost about 9–32 times as much in this study, depending on the AI. This compares average costs from the main 10,000-Creator test and the 1,000-Creator search test, not the price of every query.
- Study dates and accuracy checks
- AI queries were sent on Sep 17, 2026. Answers were checked against information from Creators’ social accounts recorded on 15–16 September 2026. We checked AI’s answers against information saved directly from Creators’ social platforms. Follower or subscriber counts could be up to 20% above or below the recorded count. Recognising the Creator alone did not make the answer accurate.
- What this does not establish
- The study does not measure lost income, verify brand partnerships, establish a wrong-person rate, identify the cause of the platform gap, or prove that Open Creator Graph improves AI answers.
Cite this study
Sparkonomy. AI Misrepresentation Study (AMISS): Findings and Methodology. September 2026.
When quoting a finding, link to this summary and include what the percentage measures.