Serious Complaint Regarding Gemini’s Repeated Hallucinations, Incorrect Accessibility Guidance, Safety Risks for Blind Users, Google AI Studio/API Problems, and the Accessibility of Google Disability Support

Gemini API
1

Dear Google Support Team,

I am writing to report a serious and ongoing problem I have experienced with Gemini while trying to use Google AI Studio and the Gemini API.

I am a blind user in South Korea. I use Windows together with Sense Reader, a screen reader, and I rely on accurate descriptions of visual interfaces and visual content when using AI systems.

I am contacting you not simply because I was unable to create a Gemini API key, but because the process exposed what I believe is a much more serious reliability and accessibility problem.

I would like to explain the situation carefully and factually so that the appropriate Google teams can investigate it.

1. The original Google AI Studio / Gemini API problem

My original goal was to obtain a Gemini API key for a personal software project.

I attempted to create a project and obtain an API key through Google AI Studio.

Instead, Google AI Studio repeatedly displayed an error similar to:

“Project creation failed. The request looks suspicious. Please try again.”

I then attempted to use Google Cloud Console as an alternative.

In the Google Cloud Console credentials/API-key interface, I encountered a situation where the Gemini API option appeared to be disabled or unavailable, accompanied by information indicating that Gemini API / Agent Platform (Vertex) required a service-account binding.

Because I use a screen reader, I depend heavily on accurate descriptions of interface controls.

Unfortunately, during the troubleshooting process, Gemini repeatedly gave me instructions that did not correspond to the interface I was actually seeing.

For example, I was instructed to select or activate controls that were not available to me. I was also told to look for options such as an “API restrictions: None” choice that I could not find in the interface.

When I explicitly explained that the control was disabled or that the alleged option was not present, the model sometimes repeated the same instruction or provided another explanation that still did not correspond to the actual interface.

This resulted in hours of repeated troubleshooting attempts.

2. The problem was not limited to one incorrect answer

What concerns me most is that this was not an isolated mistake.

The model repeatedly appeared to describe visual interface elements that were not actually present.

In several cases, I provided screenshots specifically so that the model could inspect the actual interface.

Even after seeing the screenshots, the model sometimes claimed that buttons, menus, radio buttons, options, or other controls existed when they were not present in the screenshot.

When I pointed this out, the model would sometimes change its explanation, only to provide another instruction involving an interface element that still did not exist.

This is particularly serious for a blind user.

A sighted user can independently look at the screen and immediately discover that an AI has described a nonexistent button.

A blind user using a screen reader may not have an equivalent way to independently verify a visual claim made by an AI.

As a result, an incorrect visual description can become an incorrect action.

3. Similar problems occurred with visual media

This problem has also occurred outside Google Cloud and AI Studio.

In the past, while I was using a higher-tier Gemini subscription, I showed Gemini a video that had been created using Flow and asked Gemini to examine the video.

I experienced the same general problem there.

Gemini gave descriptions and conclusions about the video that did not accurately correspond to what was actually shown.

I had to cross-check the result with other AI systems because I could not safely rely on Gemini’s description.

This was particularly disappointing because the purpose of showing the video to Gemini was precisely to have an AI system analyze visual information that I cannot independently inspect in the same way as a sighted person.

4. Repeated hallucinations and false interface descriptions

I understand that AI systems can make mistakes.

That is not, by itself, what I am reporting.

My concern is the repeated pattern:

  • The model gives an incorrect answer.
  • I provide additional information or a screenshot.
  • The model is explicitly told that the described element does not exist.
  • The model nevertheless continues to describe nonexistent controls or content.
  • It sometimes presents its new explanation with confidence rather than clearly acknowledging that it cannot verify the information.
  • I then have to repeat the process or consult another AI system to determine which information is actually correct.

This makes the problem substantially more serious than an ordinary factual error.

In my experience, the model sometimes appeared to treat an imagined or expected interface as if it were the actual interface shown in the screenshot.

For a sighted user, this may simply be frustrating.

For a blind user, it can make the system substantially less reliable because visual verification is precisely the thing the user is depending on the AI to provide.

5. The account security/recovery incident

After spending a considerable amount of time trying different troubleshooting procedures, including using an incognito browser session and attempting additional Google account verification/recovery steps, my Google account subsequently entered a security/recovery restriction.

The recovery screen stated that, if the activity had been requested by me, I should wait six hours and then check my email for a link to sign in to the account.

I want to be very careful here.

I am NOT claiming that Gemini’s incorrect instructions definitely caused Google’s security system to impose this restriction.

I do not have access to Google’s internal security logs and therefore cannot establish the causal relationship.

I am reporting the sequence of events because the restriction occurred during the same troubleshooting process.

I would like Google to determine whether the repeated account and project-creation attempts triggered any security mechanism and, if so, what actually happened.

6. Why this is particularly serious for blind users

This is the part of the problem that concerns me most.

Several days ago, I was participating in an open chat group for blind users.

I had recommended Gemini’s Live feature as a potentially useful tool for blind people, including situations such as locating an object on the floor using the phone’s camera.

I made that recommendation because I believed that a multimodal AI system could inspect the camera view and describe what it saw.

After experiencing these repeated problems myself, I now feel deeply uncomfortable about having recommended it to other blind users without first requiring much more extensive independent verification.

I feel betrayed because I trusted the system to be particularly careful when interpreting visual information.

I am especially concerned about scenarios in which a blind person uses a live camera-based AI assistant while walking outdoors and relies on the AI’s description of the surrounding environment.

If an AI system confidently describes something that is not actually present, fails to recognize an important obstacle, or incorrectly identifies the location of an object, the consequences could potentially be much more serious than simply receiving an incorrect answer.

I am not claiming that Gemini is currently unsafe for all such uses, nor am I claiming that a specific accident has occurred.

I am saying that the repeated pattern of confident visual errors raises a legitimate accessibility and safety concern that I believe Google should take seriously.

A sighted user can often independently verify an AI’s visual description.

A blind user may not have that same ability.

Therefore, the reliability standard for visual assistance can have a different practical significance for blind users.

7. The emotional impact and loss of trust

I also want to communicate the human side of this problem.

I spent hours trying to solve what initially appeared to be a straightforward API-key/project-creation problem.

Instead of receiving reliable assistance, I repeatedly received instructions that did not match the actual interface.

I then had to cross-check the answers with other AI systems.

Eventually, the troubleshooting process was followed by a Google account security/recovery restriction.

I am now at a point where I no longer know which Gemini responses I can safely trust when the task involves visual information or an unfamiliar Google interface.

For a user who is blind and specifically uses AI to compensate for visual limitations, this loss of trust is particularly serious.

8. What I am asking Google to investigate

I would like Google to investigate this matter at both the product-reliability and accessibility levels.

Specifically, I would appreciate answers to the following questions:

  1. Is Google aware of cases where Gemini incorrectly describes buttons, menus, controls, or other interface elements in screenshots even after the user points out that the described element is not present?

  2. Is Google aware of cases where Gemini incorrectly describes uploaded videos or other visual content with high confidence?

  3. Are there known limitations or known issues affecting Gemini’s visual grounding when analyzing screenshots, videos, or live camera input?

  4. Are there specific safeguards, warnings, or recommended usage practices for blind users who rely on Gemini to interpret visual environments?

  5. Does Google have a dedicated process for reporting serious visual-grounding or accessibility failures in Gemini?

  6. Can these examples be forwarded to the appropriate Gemini product/research/safety team for investigation?

  7. Regarding Google AI Studio, why was my project creation request repeatedly rejected with the message that the request appeared suspicious?

  8. Was my Google account or Google Cloud project subject to a security restriction during this process?

  9. Is the service-account binding behavior I encountered in the current Google Cloud/Gemini API interface expected under Google’s current September 2026 API system?

  10. What is the currently supported procedure for a blind user to create or obtain a Gemini API key without relying on potentially inaccurate third-party or AI-generated instructions?

  11. Are there accessibility-specific instructions for using Google AI Studio and Google Cloud with screen readers such as Sense Reader?

  12. If the Google Disability Support team is not responsible for the technical/API issue, could this report please be forwarded or referred to the appropriate Google Cloud, Gemini API, AI Studio, accessibility, or product-safety team?

9. Evidence

I can provide screenshots of the Google AI Studio error, screenshots of the Google Cloud interface, and other relevant evidence if needed.

I can also provide examples of Gemini’s incorrect descriptions and the corresponding screenshots or videos so that the responses can be compared against the actual visual content.

I would prefer that these examples be evaluated directly rather than relying solely on my description of what went wrong.

I understand that AI systems are imperfect and that a single incorrect answer does not necessarily indicate a systemic problem.

However, the repeated nature of these incidents, the fact that some errors occurred after I supplied screenshots or visual evidence, and the special consequences these errors can have for a blind user are the reasons I believe this deserves formal investigation.

I am not asking Google to simply apologize.

I am asking Google to determine what is happening, whether these behaviors are known, whether they represent a product limitation or defect, and what precautions blind users should take when relying on Gemini for visual assistance.

Thank you for taking this report seriously.

I would appreciate confirmation that the report has reached the appropriate team, or a referral to the team responsible for investigating these issues.

Sincerely,

A blind Google user in South Korea
Windows / Sense Reader user

2

Dear Google Support Team,

I am writing to report a serious and ongoing problem I have experienced with Gemini while trying to use Google AI Studio and the Gemini API.

I am a blind user in South Korea. I use Windows together with Sense Reader, a screen reader, and I rely on accurate descriptions of visual interfaces and visual content when using AI systems.

I am contacting you not simply because I was unable to create a Gemini API key, but because the process exposed what I believe is a much more serious reliability and accessibility problem.

I would like to explain the situation carefully and factually so that the appropriate Google teams can investigate it.

1. The original Google AI Studio / Gemini API problem

My original goal was to obtain a Gemini API key for a personal software project.

I attempted to create a project and obtain an API key through Google AI Studio.

Instead, Google AI Studio repeatedly displayed an error similar to:

“Project creation failed. The request looks suspicious. Please try again.”

I then attempted to use Google Cloud Console as an alternative.

In the Google Cloud Console credentials/API-key interface, I encountered a situation where the Gemini API option appeared to be disabled or unavailable, accompanied by information indicating that Gemini API / Agent Platform (Vertex) required a service-account binding.

Because I use a screen reader, I depend heavily on accurate descriptions of interface controls.

Unfortunately, during the troubleshooting process, Gemini repeatedly gave me instructions that did not correspond to the interface I was actually seeing.

For example, I was instructed to select or activate controls that were not available to me. I was also told to look for options such as an “API restrictions: None” choice that I could not find in the interface.

When I explicitly explained that the control was disabled or that the alleged option was not present, the model sometimes repeated the same instruction or provided another explanation that still did not correspond to the actual interface.

This resulted in hours of repeated troubleshooting attempts.

2. The problem was not limited to one incorrect answer

What concerns me most is that this was not an isolated mistake.

The model repeatedly appeared to describe visual interface elements that were not actually present.

In several cases, I provided screenshots specifically so that the model could inspect the actual interface.

Even after seeing the screenshots, the model sometimes claimed that buttons, menus, radio buttons, options, or other controls existed when they were not present in the screenshot.

When I pointed this out, the model would sometimes change its explanation, only to provide another instruction involving an interface element that still did not exist.

This is particularly serious for a blind user.

A sighted user can independently look at the screen and immediately discover that an AI has described a nonexistent button.

A blind user using a screen reader may not have an equivalent way to independently verify a visual claim made by an AI.

As a result, an incorrect visual description can become an incorrect action.

3. Similar problems occurred with visual media

This problem has also occurred outside Google Cloud and AI Studio.

In the past, while I was using a higher-tier Gemini subscription, I showed Gemini a video that had been created using Flow and asked Gemini to examine the video.

I experienced the same general problem there.

Gemini gave descriptions and conclusions about the video that did not accurately correspond to what was actually shown.

I had to cross-check the result with other AI systems because I could not safely rely on Gemini’s description.

This was particularly disappointing because the purpose of showing the video to Gemini was precisely to have an AI system analyze visual information that I cannot independently inspect in the same way as a sighted person.

4. Repeated hallucinations and false interface descriptions

I understand that AI systems can make mistakes.

That is not, by itself, what I am reporting.

My concern is the repeated pattern:

  • The model gives an incorrect answer.
  • I provide additional information or a screenshot.
  • The model is explicitly told that the described element does not exist.
  • The model nevertheless continues to describe nonexistent controls or content.
  • It sometimes presents its new explanation with confidence rather than clearly acknowledging that it cannot verify the information.
  • I then have to repeat the process or consult another AI system to determine which information is actually correct.

This makes the problem substantially more serious than an ordinary factual error.

In my experience, the model sometimes appeared to treat an imagined or expected interface as if it were the actual interface shown in the screenshot.

For a sighted user, this may simply be frustrating.

For a blind user, it can make the system substantially less reliable because visual verification is precisely the thing the user is depending on the AI to provide.

5. The account security/recovery incident

After spending a considerable amount of time trying different troubleshooting procedures, including using an incognito browser session and attempting additional Google account verification/recovery steps, my Google account subsequently entered a security/recovery restriction.

The recovery screen stated that, if the activity had been requested by me, I should wait six hours and then check my email for a link to sign in to the account.

I want to be very careful here.

I am NOT claiming that Gemini’s incorrect instructions definitely caused Google’s security system to impose this restriction.

I do not have access to Google’s internal security logs and therefore cannot establish the causal relationship.

I am reporting the sequence of events because the restriction occurred during the same troubleshooting process.

I would like Google to determine whether the repeated account and project-creation attempts triggered any security mechanism and, if so, what actually happened.

6. Why this is particularly serious for blind users

This is the part of the problem that concerns me most.

Several days ago, I was participating in an open chat group for blind users.

I had recommended Gemini’s Live feature as a potentially useful tool for blind people, including situations such as locating an object on the floor using the phone’s camera.

I made that recommendation because I believed that a multimodal AI system could inspect the camera view and describe what it saw.

After experiencing these repeated problems myself, I now feel deeply uncomfortable about having recommended it to other blind users without first requiring much more extensive independent verification.

I feel betrayed because I trusted the system to be particularly careful when interpreting visual information.

I am especially concerned about scenarios in which a blind person uses a live camera-based AI assistant while walking outdoors and relies on the AI’s description of the surrounding environment.

If an AI system confidently describes something that is not actually present, fails to recognize an important obstacle, or incorrectly identifies the location of an object, the consequences could potentially be much more serious than simply receiving an incorrect answer.

I am not claiming that Gemini is currently unsafe for all such uses, nor am I claiming that a specific accident has occurred.

I am saying that the repeated pattern of confident visual errors raises a legitimate accessibility and safety concern that I believe Google should take seriously.

A sighted user can often independently verify an AI’s visual description.

A blind user may not have that same ability.

Therefore, the reliability standard for visual assistance can have a different practical significance for blind users.

7. The emotional impact and loss of trust

I also want to communicate the human side of this problem.

I spent hours trying to solve what initially appeared to be a straightforward API-key/project-creation problem.

Instead of receiving reliable assistance, I repeatedly received instructions that did not match the actual interface.

I then had to cross-check the answers with other AI systems.

Eventually, the troubleshooting process was followed by a Google account security/recovery restriction.

I am now at a point where I no longer know which Gemini responses I can safely trust when the task involves visual information or an unfamiliar Google interface.

For a user who is blind and specifically uses AI to compensate for visual limitations, this loss of trust is particularly serious.

8. What I am asking Google to investigate

I would like Google to investigate this matter at both the product-reliability and accessibility levels.

Specifically, I would appreciate answers to the following questions:

  1. Is Google aware of cases where Gemini incorrectly describes buttons, menus, controls, or other interface elements in screenshots even after the user points out that the described element is not present?

  2. Is Google aware of cases where Gemini incorrectly describes uploaded videos or other visual content with high confidence?

  3. Are there known limitations or known issues affecting Gemini’s visual grounding when analyzing screenshots, videos, or live camera input?

  4. Are there specific safeguards, warnings, or recommended usage practices for blind users who rely on Gemini to interpret visual environments?

  5. Does Google have a dedicated process for reporting serious visual-grounding or accessibility failures in Gemini?

  6. Can these examples be forwarded to the appropriate Gemini product/research/safety team for investigation?

  7. Regarding Google AI Studio, why was my project creation request repeatedly rejected with the message that the request appeared suspicious?

  8. Was my Google account or Google Cloud project subject to a security restriction during this process?

  9. Is the service-account binding behavior I encountered in the current Google Cloud/Gemini API interface expected under Google’s current September 2026 API system?

  10. What is the currently supported procedure for a blind user to create or obtain a Gemini API key without relying on potentially inaccurate third-party or AI-generated instructions?

  11. Are there accessibility-specific instructions for using Google AI Studio and Google Cloud with screen readers such as Sense Reader?

  12. If the Google Disability Support team is not responsible for the technical/API issue, could this report please be forwarded or referred to the appropriate Google Cloud, Gemini API, AI Studio, accessibility, or product-safety team?

9. Evidence

I can provide screenshots of the Google AI Studio error, screenshots of the Google Cloud interface, and other relevant evidence if needed.

I can also provide examples of Gemini’s incorrect descriptions and the corresponding screenshots or videos so that the responses can be compared against the actual visual content.

I would prefer that these examples be evaluated directly rather than relying solely on my description of what went wrong.

I understand that AI systems are imperfect and that a single incorrect answer does not necessarily indicate a systemic problem.

However, the repeated nature of these incidents, the fact that some errors occurred after I supplied screenshots or visual evidence, and the special consequences these errors can have for a blind user are the reasons I believe this deserves formal investigation.

I am not asking Google to simply apologize.

I am asking Google to determine what is happening, whether these behaviors are known, whether they represent a product limitation or defect, and what precautions blind users should take when relying on Gemini for visual assistance.

Thank you for taking this report seriously.

I would appreciate confirmation that the report has reached the appropriate team, or a referral to the team responsible for investigating these issues.

Sincerely,

A blind Google user in South Korea
Windows / Sense Reader user

3

Hi everyone,

My name is Park Su-Gyo. I’m 37, blind, and a solo indie developer in South Korea. I’m not writing this as a request to Google specifically — I just wanted other developers here, including anyone who happens to work at Google, to know what I’m building and why, in case anyone has advice, feedback, or has faced similar problems.

I wasn’t always blind. I lost my eyesight later in life, and afterward had a stroke that left the right side of my body partially paralyzed. For a while I assumed games, stories, and the kind of creative work I loved were simply over for me. Two things pulled me back: accessible cinema, and the music of Andrea Bocelli, a blind singer whose work reached millions of people. He showed me that losing your sight doesn’t mean losing your ability to create something meaningful.

I’m currently building two connected projects. The larger one, Blind Man, is a big accessibility-focused RPG — audio drama plus classic RPG plus text adventure, designed so a fully blind player can complete it without ever seeing the screen. It’s a huge undertaking, so I’m not starting there.

The game I’m actually building first is called 12 Apostles. It’s a roguelike deck-building game in the style of Slay the Spire — you lead a party down through stages, building and upgrading a card deck as you fight. I chose this genre specifically because turn-based, card-driven combat with spoken menus and TTS cues can be made just as playable for a blind player as for a sighted one. 12 Apostles is my proof of concept before Blind Man’s much larger scope.

AI is central to how I can do any of this at all — for coding, debugging, TTS, translation, and multimodal review, since I can’t visually check my own game. But I don’t think AI should be given every part of the process. I believe the actual planning and design of a game has to come from a human being — that’s the only way the intent and emotion a creator wants to express actually survives into the final work. AI helps me build and realize a design once I’ve made it; the design itself stays mine.

Something I haven’t said plainly until now: I live with both blindness and a brain-based movement disorder at the same time, which means everything takes me a lot longer than it would take someone without a disability. But I intend to keep working until 12 Apostles is finished, and I believe it will turn out to be something good. I don’t know exactly how much time I have, but I plan to finish this game while I’m still alive. If I manage that, I want it to prove something bigger than just my own story: that it isn’t only blind people who can create and make art, but people with all kinds of disabilities.

(A half-serious aside: my YouTube channel is called “Blind Cat,” and given how often cats seem to end up with balance and coordination problems of their own, maybe picking up a brain-based movement disorder on top of blindness means I really have become one. Maybe becoming Blind Cat was fate all along.)

One thing I want to be honest about: I regularly see AI mistranslate or loosely paraphrase things when translating into Korean, in ways that quietly change the original meaning. This post itself was originally written in Korean and translated into English using Claude, so I genuinely don’t know how well it captures what I meant, even after translating it. As someone who can’t independently check translated text in a language I don’t speak fluently, this is a real limitation I live with.

I also want to say this directly: I want to keep trusting Gemini. Multimodal AI is part of my accessibility infrastructure, not a convenience — it’s how I understand images, video, and interfaces I can’t see myself. When it works well, it opens doors nothing else can. I’d sincerely appreciate hearing from anyone, at Google or elsewhere, who can help make that trust something I can keep relying on.

If you’re a developer, researcher, accessibility specialist, or just someone who’s tackled a project like this, I’d love your advice, feedback, or just to hear your experience.

Contact: p....@gmail.com
YouTube (dev log, Korean): 눈먼고양이 (“Blind Cat”)

Thank you for reading.
— Park Su-Gyo