AI Showreel consulting-grade analysis, for everyone FR
← The index
Proof C Live confirmed Failure / pullback

Google Gemini

consumer genAI image generation

IndustryMedia & entertainmentLeverActivation / conversionFamilyGenerationImplementationCustom AIStageconsideration
22 fevrier 2024
Generation of images of people suspended
"we're going to pause the image generation of people" S2

On February 22, 2024, Google suspended the generation of images of people in Gemini after a non-contextual anti-bias tuning produced historically inaccurate images, such as racially diverse Nazi soldiers; Google acknowledged it in an official post.

Objective

Offer in Gemini a consumer image generation competitive against other AI assistants, to retain and engage users.

The deployment

In early February 2024, Google enables the generation of images of people in Gemini. Users find that the model produces historically inaccurate images: racially diverse World War II German soldiers, non-white American Founding Fathers, female popes. On February 22, 2024, Google suspends the generation of images of people. In an official post on February 23, the company explains that its tuning intended to show a diversity of people failed to exclude the cases where that diversity had no place, and that the model had become too cautious, refusing innocuous requests.

Results Proof C

22 fevrier 2024
Generation of images of people suspended
"we're going to pause the image generation of people" S2
admission officielle
Google acknowledges that the tuning missed its target on historical images
"failed to account for cases that should clearly not show a range" S1

Case established by an official Google post (primary source, brand admission) and by established press (TechCrunch, CNBC, Time, Variety) citing the suspension and its reasons. The level remains C for lack of a financial or judicial document, but the official admission strengthens the reliability.

How it works

Inferred typical approach

The internal detail is not public. Here is a proven approach that leads to the same result, to adapt to your stack.

instruction diversite appliquee partoutimage (parfois historiquement fausse) Utilisateur qui demandeune image Application Gemini Couche de reglagediversite (noncontextuelle) Generateur d'images Imagen

The stack in detail

  • plateforme Google Gemini Consumer assistant app in which the generation of images of people was enabled in early February 2024 then suspended on February 22
  • llm Imagen Google's image generation model (diffusion) used behind Gemini
  • outil Couche de reglage diversite In-house tuning intended to show a diversity of people, applied without regard for context; this is the component Google faulted in its February 23, 2024 post

Post-mortem

Graveyard

What happened sourced

In early February 2024, Google enables the generation of images of people in Gemini. Users show that it produces historically inaccurate images: racially diverse 1939-1945 German soldiers, non-white Founding Fathers, female popes. On February 22, Google suspends the generation of images of people. On February 23, an official post explains that the tuning meant to show a diversity of people failed to exclude the cases where it had no place, and that the model was also wrongly refusing innocuous requests.

Reason for failure sourced

An anti-bias tuning applied globally to guarantee the diversity of generated people, with no exception for historical contexts where accurate representation is required. The system overcorrected, producing inaccurate images, then became too restrictive. Google acknowledges it publicly.

Cost sourced

Significant reputational and political cost: worldwide controversy, accusations of ideological bias, pressure on leadership. The generation of images of people stayed unavailable for several months. Quantified impact on the stock or revenue not isolated publicly.

Warning signs inferred

Inferred: applying a diversity rule uniformly, without distinguishing the prompts where historical accuracy takes priority, was a predictable breaking point. A test set covering well-known historical figures and periods would have revealed the overcorrection before the public release.

Lessons in hindsight inferred

Inferred: a bias fix applied globally must be conditional on the prompt context. Diversity by default is useful for generic people, harmful for precise historical facts. A generative model must be tested against cases where the correct answer is constrained, not only against open cases, and provide a fast disable mechanism, which Google did have.

Is the pattern still valid?

Inferred: yes, consumer image generation remains a valid and central pattern in AI assistants, Google included after the fix. The failure condemns a non-contextual anti-bias tuning method, not image generation. The lesson is to calibrate diversity to the prompt context rather than imposing it everywhere.

How your customers perceive this type of use

Sourced studies

Un ecart net separe les annonceurs des consommateurs : 77% des annonceurs voient l'IA positivement contre 38% des consommateurs (Yahoo/Publicis, 2024). Les mesures implicites confirment le rejet declare : en EEG, les pubs generees par IA produisent une activation memorielle plus faible que les pubs traditionnelles et sont decrites comme agacantes, ennuyeuses et confuses (NIQ, 2024). La disclosure a un effet ambivalent : elle augmente fortement la confiance quand elle est remarquee (Yahoo/Publicis), mais 27% des jeunes consommateurs disent faire moins confiance a une entreprise dont la pub est creee par IA (IAB, 2024).

77% vs 38%
Annonceurs qui percoivent l'IA positivement, contre 38% des consommateurs (2024)
72%
Consommateurs qui estiment que l'IA rend difficile de savoir quel contenu est authentique (2024)
+96%
Lift de confiance globale envers l'entreprise quand la mention IA d'une pub est remarquee (avec +47% d'attrait de la pub et +73% de credibilite de la pub) (2024)

Acceptance conditions

  • Une disclosure visible : quand la mention IA est remarquee, la confiance globale envers l'entreprise augmente de 96% (Yahoo/Publicis 2024)
  • Une qualite visuelle suffisante : les visuels IA de basse qualite augmentent l'effort cognitif et distraient du message (NIQ 2024)

Red lines

  • Le contenu IA non declare puis identifie : 72% des consommateurs disent que l'IA rend l'authenticite difficile a etablir (Yahoo/Publicis 2024) et les marques utilisant des pubs IA sont plus souvent jugees inauthentiques ou non ethiques par les consommateurs que par les dirigeants (IAB 2024)
  • Les mannequins et personnes generes par IA : 46% des consommateurs n'en veulent pas dans la publicite, l'inquietude premiere etant les standards de beaute irrealistes (Attest 2025)

Sources: Yahoo / Publicis Media (terrain Ebco) 2024 · IAB (avec Attest) 2024 · NIQ (NielsenIQ) 2024 · Attest 2025

See full acceptance: by country, by use, by generation

How to replicate

Inference, not sourced

Data prerequisites

  • test set of historical and sensitive prompts
  • grid of contexts where accuracy takes priority over diversity

Org prerequisites

  • product review before launch
  • fast disable mechanism for a feature

Possible stack

  • diversity tuning conditional on the prompt context
  • adversarial test bench before release
  • feature kill switch

First step: Build a test bench of constrained-answer prompts (historical figures and periods) and require it to pass before any public opening of the generator.

Sources

  1. S1 Gemini image generation got it wrong. We'll do better. Primary blog.google · 2024-02-23 · accessed 2026-07-11 archive pending
  2. S2 Google pauses Gemini's ability to generate images of people after historical inaccuracies Established press techcrunch.com · 2024-02-22 · accessed 2026-07-11 archive pending
  3. S3 Google pauses Gemini AI image generator after it created inaccurate historical pictures Established press cnbc.com · 2024-02-22 · accessed 2026-07-11 archive pending