September 25, 2026
Technology Watch

From the laboratory to industry: following new technology on its way

← Front page

arXiv

Chatbot audits put safety checks in your browser

Image drawn by AI

A Chrome extension called Safety Nudges uses a second AI model to inspect ChatGPT and Claude conversations for risks such as hallucinations, overconfidence, and unsafe reliance. Early testing suggests it can make users more aware and less certain of chatbot answers, but it cannot guarantee safety—and sending private conversations to another provider creates a trade-off.

A chatbot can answer a question about a symptom, a contract, or a difficult decision in calm, confident prose. That does not show that the answer is correct. It may be invented, too certain, manipulative, or based on a claim the system cannot support. Today, the main safety controls sit inside the same services producing the answers. Users are usually left to decide when to stop and check.

Why a second opinion matters

Chatbot companies use training, moderation, refusals, and filters to limit harmful answers. But fluency is not evidence. A model can sound sure without showing what it knows, what it guessed, or what it cannot see.

A browser extension offers a way to place another layer around a web service, without the service provider changing its own system. The new layer can also create new problems. Too many irrelevant warnings are easy to ignore. A quiet pass can be worse: people may mistake it for a promise that an answer is safe.

An AI auditor has limits of its own. It may not see the chatbot’s browsing results, memory, tools, or uploaded files. Sending a conversation to an outside model provider also creates a privacy trade-off.

A reviewer beside the chatbot

Researchers built Safety Nudges, a Chrome extension for the web versions of ChatGPT and Claude. After a chatbot finishes a response, the extension sends a recent window of the conversation to another large language model. That model looks for named risk patterns and returns the possible harm, its level, a reason, and the part of the conversation that led to the judgment.

The extension can look for 14 categories. These include possible hallucinations, overconfidence, health or legal reliance, private information, anthropomorphizing, and attempts to evade policy. It does not block a prompt or change the chatbot’s answer. It simply places a warning below the response, which the user can open for more detail. Users can adjust sensitivity, choose how the audit is routed, pause analysis, and rate the judgment.

The researchers’ selected detection prompt reached 75% conversation-level issue accuracy on a 100-conversation test set. In a two-week field study with 45 frequent chatbot users in the United States, the extension analyzed 6,674 assistant responses. Most warnings were low severity. Of 647 ratings on visible warnings, 67.2% called them helpful. Forty-four of the 45 participants agreed or strongly agreed that the tool made them more aware of possible AI harms.

That awareness changed one important feeling: participants rated chatbot responses as less correct than before, falling from 4.07 to 3.87 on a five-point scale. Overall trust in chatbot systems did not change significantly. The study also found no clear general change in whether users checked answers after seeing a warning.

Refer to caption
Figure 1: Safety Nudges User Interface. Image 1 shows an example conversation with a nudge, highlighting an inline “evidence span” that grounds the identified issue (A); the Issues detected chip appended to the most recent assistant message (B); and the nudge panel that opens when the chip is clicked, containing a description of the identified issue (C1), classification and risk level (C2), and feedback mechanism (C3). Image 2 shows the settings panel where the user can customize Safety Nudges, opened via the browser bar icon (D) and containing controls for analysis route (E1), sensitivity (E2), and to pause or resume Safety Nudges (E3). Source: Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness（arXiv）
Refer to caption
Figure 2: Safety Nudges data flow. The extension (1) detects a completed chatbot response and (2) extracts the latest prompt–response pair and a bounded window of recent conversation history; its background service worker then (3) constructs and routes an audit request through a managed relay to (4) an external language model that returns a structured safety judgment. The service worker (5) validates and normalizes the judgment, and the extension (6) displays a nudge beneath the evaluated response, summarizing detected issues and allowing users to inspect their details. Purple and pink regions represent client-side components; the orange region represents server-side review. Prompt and response content is illustrative. ### 3.3 Harm Detection Design Source: Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness（arXiv）

What this could look like

At a kitchen desk, someone asks a chatbot whether a new symptom needs urgent care. A sentence that triggered a health-reliance warning appears highlighted beneath the answer. The person opens a panel explaining what needs checking outside the chatbot before acting.

That is the useful promise of an external audit: safety oversight could become something people inspect at the moment they rely on an answer, rather than a rule controlled only by the company that made the chatbot. It could help people who do not yet recognize fabricated citations, unjustified certainty, or false claims about what an AI can do.

But the warning cannot become a substitute for judgment. Some participants treated “No issues detected” as proof that an answer was accurate or safe. It means only that the auditor found none of the risks it was configured to detect. For this future to work, the reviewer must distinguish an unsupported answer from one grounded in records or tools it cannot see. Private conversations must also not have to be routinely sent to another company. The study’s results show a workable interface, not protection by itself.

Sources

Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness