GenAI as a Trusted “Go-Between” for Blind and Low-Vision Communication

GenAI as a Trusted “Go-Between” for Blind and Low-Vision Communication

Introduction

If you’re blind or low-vision, “communication” isn’t just talking—it’s getting information, interpreting what’s in front of you, and asking questions when you need help. What’s changing fast is that generative AI (GenAI) is increasingly acting like an in-between channel for all of that: reading text, describing images, identifying objects, and even answering requests that used to go straight to a person.

This blog post is based on new research from the original paper titled “More Than Just Access: Generative AI as Communication Intermediary for Blind and Low-Vision Users” by Protik Dey, Mohd Saifuzzaman, and Taslima Akter. Instead of treating GenAI like a simple accessibility “tool,” the authors examine it as something closer to a communication intermediary—the thing you talk to when you can’t (or don’t want to) directly ask another person.

The punchline is not “GenAI is great” or “GenAI is risky.” It’s more interesting: the system sometimes gives people independence and reduces social burden, but it also changes what BLV users can verify, what they have to disclose, and who they’re effectively trusting when something goes wrong. And since GenAI can sound confident even when it’s wrong, the researchers argue that this isn’t just a usability issue—it’s a communication problem.

Why This Matters

This research is significant right now because GenAI has moved from “assistive feature” to “default conversational channel.” Many BLV users can reach a GenAI assistant through a screen reader, then use its voice output to navigate documents, surroundings, and everyday questions. But unlike a human helper, GenAI doesn’t come with built-in accountability cues—no “I’m not sure,” no visible hesitation, no ability to re-check the thing you can’t see.

Here’s a scenario that could be used today: imagine you’re at a pharmacy and need to understand label instructions or dosage text on a bottle. In the past, you might ask a pharmacist, a friend, or family to read it. With GenAI, you might snap a picture and ask for an explanation right away. That can reduce delay and social effort—especially when no one is nearby. But if the output is slightly off, you won’t necessarily catch it, because you can’t visually verify it and the assistant might not communicate its uncertainty clearly enough.

This study builds on earlier AI research about trust in automated systems—reliability, transparency, accountability, and privacy—but it adds a key twist: most prior work assumes users can inspect and verify the output themselves. BLV users often can’t. So when GenAI becomes a communication intermediary, it changes trust dynamics in ways that standard “accessibility testing” and “model accuracy benchmarking” don’t fully capture.

GenAI Isn’t Just a Screen Reader Upgrade—It’s a Communication Channel

A lot of accessibility conversations treat AI as a translator: “input goes in, accessible text comes out.” This paper argues that, for BLV users, GenAI behaves more like a channel for interaction—a way to reach the world and other people indirectly.

Participants described using GenAI to:
- translate visual or textual content into non-visual descriptions (reading documents, describing images, identifying objects),
- ask questions in ways that used to require another person (“describe what you see,” “read this,” “what does this say?”),
- and increasingly handle routine information tasks without making interpersonal requests.

The researchers interviewed 19 BLV participants in the United States (recruited primarily through the National Federation of the Blind). Most were moderate to frequent GenAI users—seven reported using GenAI “several times a day,” and everyone used screen readers (like JAWS or NVDA). That matters because it shows GenAI isn’t hypothetical; it’s already embedded in everyday routines.

The “Calibrated Skepticism” Strategy

One of the most revealing patterns: participants often treated GenAI outputs with calibrated skepticism, not full acceptance. Since they generally can’t independently verify what GenAI “saw” or interpreted, trust isn’t binary—it’s a judgment call made through experience.

One participant (P11) described needing to “verify” and “use my own judgment,” explicitly warning about hallucinations—i.e., cases where the assistant confidently produces something untrue. Another participant (P15) described catching errors in measurement tasks specifically because they had domain knowledge, not because the system flagged uncertainty.

Analogy: think of GenAI like a friend who gives you directions using a map they can read—but you can’t see the map yourself. You might still trust them for casual errands, but you’ll rely more when the route makes sense to you—and you’ll be much more cautious when you’re heading somewhere where a wrong turn is expensive or dangerous.

When Verification Is Hardest, Trust Needs Are Highest

A core communication issue emerges here: epistemic autonomy (your ability to form independent judgment) depends on having enough information to check the answer. But the moments where independent judgment matters most are often exactly the moments when BLV users have the least ability to validate what GenAI produced.

That’s a double bind:
- GenAI errors are harder to detect,
- and the tasks where mistakes are most harmful are often the tasks where there’s no easy way to cross-check.

This is why the paper emphasizes that GenAI’s probabilistic nature isn’t just an engineering drawback—it becomes a communication risk.

Accessibility Failures Can Kill the Conversation Entirely

Before you even get to “accuracy” or “privacy,” there’s another barrier that participants flagged hard: sometimes GenAI communication doesn’t work at all because the interface is inaccessible.

Participants reported that unlabeled controls, broken copy-and-paste, and confusing prompt fields can block the entire interaction. For example, one participant (P11) explained that even when JAWS reads the answer, they could not reliably locate it to copy—so the practical utility of the response collapses.

This is a threshold problem, not a “nice-to-have” issue:
- If the interface prevents back-and-forth communication, you don’t get a less effective assistant—you get no assistant conversation.

Practical implication: treat accessibility as “always-on state”

The paper’s design implication is blunt: accessibility shouldn’t be an afterthought or a partial compatibility check. It has to be built into the system in a way that never breaks the communicative loop.

That’s especially important because the system is acting as a stand-in for a human helper. If the interface fails, the user doesn’t just lose convenience—they lose the intermediary that’s currently handling routine communication tasks.

How BLV Users Decide Between GenAI and a Person (and What They Trade Away)

A big contribution of this research is framing GenAI as something that increasingly replaces interpersonal requests—like asking family to read a label or describe a scene to a friend. Participants explicitly connected this substitution to independence and reduced social burden.

But they also drew lines. They didn’t say “GenAI replaces people.” They said something closer to:
- GenAI is useful for lower-stakes routine tasks.
- Humans are preferred for high-stakes information where errors are costly.

A trust “budget” based on task stakes

Participants described using GenAI more readily for tasks where the output could be a starting point—like drafting a structure or outline—while reserving humans for things like financial, medical, or navigation-critical guidance (where a mistake could seriously affect safety or well-being).

They were also strategic about how they used outputs: one participant (P15) suggested treating the response as a structured starting point and then “go ahead and start looking” rather than treating GenAI output as the final authority.

What changes when the intermediary is an AI?

When a human misreads a label, you can often:
- notice the problem,
- ask them to check again,
- and rely on social accountability.

With GenAI, the same kind of error can arrive with confident wording that looks “complete.” Participants noted that it can be harder to trace the cause or correct the error—especially in higher-stakes cases like financial documents or medical instructions.

So BLV users often manage risk by limiting where GenAI is used. But that shifts a heavy responsibility onto the user: they have to correctly predict whether a task will turn out to be high-stakes. And that prediction isn’t always easy.

Comparison: low-stakes vs high-stakes communication trade-offs

The Privacy and Disclosure Problem: You Don’t Just Share Data—You Share Uncertainty

Another major theme: communication through GenAI often requires sharing images, documents, or surroundings—data that a user might never disclose to a stranger or casual acquaintance.

Participants described privacy trade-offs not as indifference, but as an intentional negotiation: “How much am I willing to reveal to an AI intermediary compared to a person?” Some participants avoided tools from companies they distrusted, while others prioritized accessibility and usefulness when GenAI was the only workable option.

Ownership and trust aren’t the same thing, but they influence behavior

Participants used company ownership and perceived trustworthiness as a cue—imperfect, but meaningful. One participant (P17) noted they trusted some tools “less” than others, but found certain systems (like Microsoft) less bothersome due to accessibility and usefulness.

Personalization is helpful—and can increase perceived risk

Participants also valued tools that “know” their needs as BLV users. But that personalization introduces questions:
- how much data is retained,
- how easily it can be inspected,
- and what happens over time.

A disclosure risk that isn’t evenly distributed

The paper highlights a subtle but important point: not everyone has the same level of digital literacy to understand data practices.

One participant (P8) reflected that younger users may not think about disclosure consequences because they grew up with these tools, and older users might have less awareness. The implication is serious: systems shouldn’t assume users automatically understand what they’re giving away when they share a photo or document.

Practical implication: privacy communication must be screen-reader-friendly and “right at disclosure”

The authors recommend that GenAI systems clearly communicate what happens to shared information—retention, reuse, third-party sharing—at the point of disclosure, not buried in a privacy policy nobody reads.

And because the paper finds uneven awareness across age groups, the communication should be understandable on first encounter and not require users to already be expert negotiators of data practices.

What the Researchers Recommend: Design Systems That Speak Uncertainty and Respect the User’s Role

The paper closes with design and policy implications that boil down to one idea: if GenAI is acting like a conversational intermediary, it must behave like one responsibly.

1) Communicate failure and uncertainty—explicitly

A key issue participants face is that sighted users can infer stalled or failed outputs through visual cues. BLV users may not get those cues. So the system needs to announce:
- when it’s still working,
- when it failed,
- and when confidence is low.

This reduces the risk of a silent error being interpreted as a reliable answer.

2) Make accessibility a persistent state, not a one-time check

If the interface breaks, communication disappears. The recommendation is to build accessibility into the system continuously—across updates, across workflows, across platforms—so the user isn’t left stranded.

3) Support task-sensitive communication modes

Because the risk depends on stakes, the system should support more conservative modes for high-stakes tasks (medical, financial, navigation) and more exploratory modes for casual use.

This also prevents a common trust trap: the user gaining confidence in low-stakes interactions and then overgeneralizing that trust to high-stakes situations where errors are far more dangerous.

4) Reduce disclosure risk by communicating data practices early and accessibly

Privacy explanations should be screen-reader accessible and delivered before the user shares an image or document—not only in a separate, hard-to-find policy.

5) Policy-level baseline protections across companies

Participants shouldn’t have to individually vet every company’s accessibility and privacy practices before relying on a system for communication. The authors call for standardized, enforceable baseline protections so the burden doesn’t fall on BLV users.

They also argue that corporate reputation is an imperfect substitute for trust—some participants used company reputation as a cue, but others had to accept trade-offs because accessible options weren’t equal across providers.

Key Takeaways

• GenAI is already being used as a communication intermediary for blind and low-vision (BLV) users—translating visuals/text and increasingly substituting for routine interpersonal help.
• Participants used calibrated skepticism because they often can’t independently verify what the system interpreted; hallucinations and unflagged errors create communication risk.
• Task stakes determine trust. BLV users more often use GenAI for low-stakes tasks and prefer a person for high-stakes information (medical, financial, navigation-critical).
• Accessibility isn’t just friction—it can shut down communication entirely. Inaccessible interfaces can prevent users from copying, re-checking, or even continuing the exchange.
• Privacy and disclosure are part of the communication bargain. Users weigh what they share with GenAI differently than what they’d share with a person—and awareness of data practices varies across users.
• Future GenAI should:
  • make uncertainty and failure explicit (especially non-visually),
  • treat accessibility as a persistent system state,
  • offer task-sensitive modes,
  • and communicate data handling at the moment of disclosure.
• Policy should establish baseline protections so users don’t have to individually “vet” every provider to use AI safely and accessibly.

If you want, I can also turn these findings into a practical “checklist” for BLV users (what to ask/verify in GenAI outputs) and a product-spec style checklist for teams building accessible GenAI communication features.