Here is what the proposed federal Stealth Bot Prohibition Act, H.R. 9915 prohibits:

A person may not (1) deploy, direct, or cause to be deployed, a stealth bot to access a website, digital platform, or online service in a manner that is reasonably likely to damage, impair, or burden the technical or commercial operation of a website or other internet service; or (2) intentionally misrepresent, conceal, or obscure the nature of the bot to appear as a human user for use in connection with a generative AI model or service.

The statute starts from the premise that a bot visiting a website should not be permitted to masquerade as an ordinary human visitor. But an interesting secondary question arises by considering what should happen when anonymity is not evidence of bad behavior, but part of what makes an automated tool (such as an AI agent visiting a website) useful.

Journalists use automated tools to investigate websites without tipping off their subjects. Researchers collect information to study discrimination and pricing practices. Cybersecurity researchers look for vulnerabilities. Companies monitor competitors and compare prices. Archives preserve material before it disappears. Lawyers visit sites to gather evidence for use in enforcement and dispute resolution. And increasingly, AI agents visit websites on behalf of individual users.

That makes the Stealth Bot Prohibition Act potentially more consequential than a law directed only at large AI companies scraping the web for training data.

What is a stealth bot?

Representatives Laurel Lee, Valerie Foushee and Gus Bilirakis introduced H.R. 9915 on July 23, 2026. The bill was referred to the House Committee on Energy and Commerce, which remains the last action listed by GovInfo as of the date of this article.

The legislation starts with an expansive definition of a “bot.” It means software that “retrieves, scans, indexes, scrapes, or otherwise accesses” a website or other internet source. The definition specifically encompasses crawlers, spiders, fetchers, clients, user agents and AI agents.

A “stealth bot” is one that interacts with a website, digital platform or online service without prior disclosure of its identity and purpose. Among other things, the bill points to failure to use a valid and accurate user-agent string and failure to disclose the bot’s specific nature and purpose.

And the contemplated disclosure goes well beyond simply announcing, “I am a bot.” The bill would call for disclosure of proposed uses of page content, expressly mentioning text and data mining, search indexing, inferencing, AI development and operations, training, fine tuning and retrieval-augmented generation. That disclosure is supposed to occur when access is requested and in a format accessible to the website operator.

That presents a practical question right away: how much must a machine reveal about itself before it knocks on the door?

The bill does not simply ban every unidentified bot

There is an important qualification in the operative language. The bill would prohibit a person from deploying, directing or causing a stealth bot to access a website or online service when the manner of access is reasonably likely to “damage, impair, or burden” its technical or commercial operation. Separately, it would prohibit intentionally disguising the nature of a bot to make it appear human when the bot is used in connection with a generative AI model or service.

Those are distinct prohibitions. The first combines lack of disclosure with a specified type of potential harm. The second addresses intentional human impersonation in connection with generative AI without using the same “damage, impair, or burden” language.

That distinction matters because much of the public debate around stealth bots uses the term more broadly than the operative federal prohibition does.

Why would a legitimate bot want to stay hidden?

Kevin Bankston of the Center for Democracy & Technology has written about the dispute as part of a much larger struggle over the future of automated access to the open web. He points out that bots have long performed functions such as search indexing, security testing, accessibility auditing, comparison shopping, academic research, RSS retrieval and web archiving. AI has dramatically increased the stakes, he argues, but automated access itself is hardly new.

Some automated activities may work differently when the website knows precisely who is conducting them. Consider an investigation into personalized prices. A researcher might send automated requests from different locations or under different circumstances to determine whether a company is showing different users different prices. If every request announces the researcher, the organization conducting the study and the purpose of the experiment, the observation itself could affect the behavior being observed.

Re:Create has raised this concern specifically in connection with the federal legislation, arguing that some journalists need automated tools that do not announce the investigation to its subject. Its July analysis contends that mandatory identification can interfere with reporting techniques used to investigate matters such as dynamic pricing and algorithmic discrimination.

There is an obvious countervailing concern. A publisher receiving millions of automated requests may have no easy way to determine whether those requests come from a search engine, an AI company, an intermediary gathering data for resale, a researcher or something else entirely.

Sara Guaglione described that problem in Digiday. Her reporting explains some of the techniques used to make automated collection resemble ordinary human browsing, including generic browser user agents and residential IP addresses. From the publisher’s perspective, inability to identify the crawler can make it difficult both to manage server traffic and to determine who is collecting content and for what purpose.

So the hard question is not whether transparency is generally good. It is whether every useful form of automated access can operate effectively under the same transparency rule.

This could matter to companies that have nothing to do with training an LLM

The text of H.R. 9915 does not define a bot as something used only to train an artificial intelligence model. Its definition reaches software that retrieves, scans, indexes, scrapes or otherwise accesses online sources. That potentially puts a much broader range of business activity into the conversation.

A retailer may collect publicly available prices from competitors. A brand may hire a vendor to monitor unauthorized sales or counterfeit products. A financial-services company may purchase web-derived data. A cybersecurity company may scan sites. A travel service may aggregate publicly available information. A company may use an AI agent to navigate websites and retrieve information for employees or customers. None of those activities necessarily involves building a large language model.

The bill also applies to a person who does not personally deploy the bot but “direct[s]” or “cause[s]” one to be deployed. That language creates an important business question around outsourced data collection. And it also is an interesting attempt to address, by means of codified law, some of the issues we are seeing around the behavior of agentic AI systems. Think of the Amazon v. Perplexity situation, for example.

A company that hires a vendor to gather information may therefore want to know more than whether the vendor’s dataset was obtained from “publicly available sources.” How did the vendor collect it? What did its automated tools identify themselves as? What purpose did they disclose? Did they pose as human visitors? And how much direction did the customer give about where and how the information should be obtained? The proposed law could thus become relevant not only to crawler developers but also to vendor diligence and technology contracting.

What exactly must the bot disclose?

The mechanics may prove just as important as the principle. The bill specifically mentions a valid and accurate user-agent string, but that is only one part of the definition. It separately calls for disclosure of the bot’s “specific nature and purpose,” including proposed uses of page content.

Suppose a company identifies its software accurately as AcmeBot. Is “commercial research” enough to describe its purpose? What about “AI purposes”? If the company may use the information both for search and retrieval-augmented generation, must it disclose both?

And what happens when the use changes later? A dataset collected for search today might become useful for an AI application next year. A corpus gathered for analytics could later be used for fine tuning. Information acquired by one business unit could become useful to another.

Because the bill speaks in terms of a “proposed use and purpose” when access is requested, companies could eventually face not only a provenance question, Where did this data come from?, but also a purpose-provenance question: What did we say we were going to do with it when we collected it?  That could make collection history more important to the future usability of data.

And what does “commercial burden” mean?

The first prohibition applies when a stealth bot accesses an online service in a manner reasonably likely to “damage, impair, or burden” its technical or commercial operation. H.R. 9915 does not define those concepts further.

There is something familiar about that wording for lawyers who have spent time with the Computer Fraud and Abuse Act. The CFAA has generated years of litigation over the distinction between harm to a computer system itself and the economic consequences flowing from the conduct. Under that statute, “damage” and “loss” are separately defined concepts, and a business cannot necessarily turn every economic injury associated with computer access into the type of technological harm the statute recognizes. An internetcases post discussing Cassetica Software, Inc. v. Computer Sciences Corp., for example, examined a court’s conclusion that unauthorized software downloads did not establish CFAA “damage,” while lost licensing fees did not qualify as statutory “loss” because they were not tied to an interruption of service.

H.R. 9915 uses different language and creates a different statutory scheme, so the CFAA cases would not supply a ready-made interpretation. But the resemblance points to the same basic line-drawing problem. A bot that overwhelms servers or makes a service unavailable presents an obvious case of burdening technical operations. It is less obvious what Congress means by burdening “commercial” operations.

Would that include increased infrastructure expenses? Lost advertising impressions because an AI service answers a question without sending the user to the source? Reduced subscription revenue? Lost licensing opportunities? Competitive harm from another company incorporating scraped information into its own product?

The bill does not say. If the legislation becomes law in its present form, the meaning of “commercial operation” could therefore become an important boundary question, particularly because it may determine when an undisclosed bot crosses the line from merely being difficult to identify to engaging in prohibited conduct.

Does disclosure mean permission?

No provision in the introduced bill says that identifying a bot gives it authorization to take or use whatever it encounters. In fact, H.R. 9915 expressly says that its provisions do not affect rights, limitations or remedies available under other federal or state law or the Constitution.

So a bot could conceivably comply with the disclosure requirements while leaving entirely separate disputes about copyright, contract, privacy, computer access or other legal restrictions. That is an important distinction for both sides of an automated transaction: identification is not the same thing as authorization.

The broader question is what kind of web we want machines to see

The immediate policy dispute concerns AI companies, publishers and automated scraping. But the deeper issue is older than generative AI. The web was built in an environment in which people and machines could visit publicly available pages without first establishing who they were. Search engines, archives, researchers and innumerable other automated services grew up inside that architecture.

Generative AI has put enormous pressure on that arrangement. Publishers have reasons to want to know when automated systems are consuming their content, imposing infrastructure costs or using material to create competing products. At the same time, journalists, researchers and other actors may have reasons not to disclose themselves before observing a public-facing system. Those competing concerns are reflected in the sharply different perspectives speaking on the debate.

H.R. 9915 therefore raises a question considerably broader than whether OpenAI, Google or another large AI company should identify its crawler. It asks what information any automated visitor should have to reveal before accessing the open web, and whether the answer should be the same for every bot.

For businesses, that makes the proposal worth watching from both directions. If the bill passes, companies that operate websites may get more information about who is sending automated traffic to their sites and what those bots intend to do with the content they collect. Companies that use bots, AI agents, data vendors or automated research tools may need to understand what those systems disclose about themselves before they collect information.

See also this circa 2020 explanation of web scraping: