#artificial-intelligence learning platform

Carroll says he only discovered that appalling lack of security around applicants' information because he was intrigued by McDonald's decision to subject potential new hires to an AI chatbot screener and personality test. “I just thought it was pretty uniquely dystopian compared to a normal hiring process, right? And that's what made me want to look into it more,” says Carroll. “So I started applying for a job, and then after 30 minutes, we had full access to virtually every application that's ever been made to McDonald's going back years.”

When WIRED reached out to McDonald’s and Paradox.ai for comment, a spokesperson for Paradox.ai shared a blog post the company planned to publish that confirmed Carroll and Curry’s findings. The company noted that only a fraction of the records Carroll and Curry accessed contained personal information, and said it had verified that the administrator account with the “123456” password that exposed the information “was not accessed by any third party” other than the researchers. The company also added that it’s instituting a bug bounty program to better catch security vulnerabilities in the future. “We do not take this matter lightly, even though it was resolved swiftly and effectively,” Paradox.ai’s chief legal officer, Stephanie King, told WIRED in an interview. “We own this.”

In its own statement to WIRED, McDonald’s agreed that Paradox.ai was to blame. “We’re disappointed by this unacceptable vulnerability from a third-party provider, Paradox.ai. As soon as we learned of the issue, we mandated Paradox.ai to remediate the issue immediately, and it was resolved on the same day it was reported to us,” the statement reads. “We take our commitment to cyber security seriously and will continue to hold our third-party providers accountable to meeting our standards of data protection.”

Carroll says he became interested in the security of the McHire website after spotting a Reddit post complaining about McDonald's hiring chatbot wasting applicants' time with nonsense responses and misunderstandings. He and Curry started talking to the chatbot themselves, testing it for “prompt injection” vulnerabilities that can enable someone to hijack a large language model and bypass its safeguards by sending it certain commands. When they couldn't find any such flaws, they decided to see what would happen if they signed up as a McDonald's franchisee to get access to the backend of the site, but instead spotted a curious login link on McHire.com for staff at Paradox.ai, the company that built the site.

On a whim, Carroll says he tried two of the most common sets of login credentials: The username and password “admin," and then the username and password “123456.” The second of those two tries worked. “It's more common than you'd think,” Carroll says. There appeared to be no multifactor authentication for that Paradox.ai login page.

With those credentials, Carroll and Curry could see they now had administrator access to a test McDonald's “restaurant” on McHire, and they figured out all the employees listed there appeared to be Paradox.ai developers, seemingly based in Vietnam. They found a link within the platform to apparent test job postings for that nonexistent McDonald's location, clicked on one posting, applied to it, and could see their own application on the backend system they now had access to. (In its blog post, Paradox.ai notes that the test account had “not been logged into since 2019 and frankly, should have been decommissioned.”)

That's when Carroll and Curry discovered the second critical vulnerability in McHire: When they started messing with the applicant ID number for their application—a number somewhere above 64 million—they found that they could increment it down to a smaller number and see someone else's chat logs and contact information.

The two security researchers hesitated to access too many applicants' records for fear of privacy violations or hacking charges, but when they spot-checked a handful of the 64-million-plus IDs, all of them showed very real applicant information. (Paradox.ai says that the researchers accessed seven records in total, and five contained personal information of people who had interacted with the McHire site.) Carroll and Curry also shared with WIRED a small sample of the applicants' names, contact information, and the date of their applications. WIRED got in touch with two applicants via their exposed contact information, and they confirmed they had applied for jobs at McDonald's on the specified dates.

The personal information exposed by Paradox.ai's security lapses isn't the most sensitive, Carroll and Curry note. But the risk for the applicants, they argue, was heightened by the fact that the data is associated with the knowledge of their employment at McDonald's—or their intention to get a job there. “Had someone exploited this, the phishing risk would have actually been massive,” says Curry. “It's not just people's personally identifiable information and résumé. It's that information for people who are looking for a job at McDonald's, people who are eager and waiting for emails back.”

That means the data could have been used by fraudsters impersonating McDonald's recruiters and asking for financial information to set up a direct deposit, for instance. “If you wanted to do some sort of payroll scam, this is a good approach,” Curry says.

The exposure of applicants' attempts—and in some cases failures—to get what is often a minimum-wage job could also be a source of embarrassment, the two hackers point out. But Carroll notes that he would never suggest that anyone should be ashamed of working under the Golden Arches.

AI and AI Detectors

The term AI detector is fundamentally misleading. These tools do not detect artificial intelligence in the way a scanner detects radiation or a forensic test detects DNA. They do not identify watermarks, embedded signatures, or any hidden metadata left behind by generative models. No contemporary AIs (including ChatGPT, Gemini, Claude, or open-source equivalents) inserts any such traceable markers into its text output. Once generated, the text exists as ordinary language, indistinguishable in form from human-authored writing.

Because there is nothing intrinsic to detect, so-called AI detectors instead rely on statistical inference. They are themselves AIs trained on large corpora of text labeled as “human-written” or “AI-generated.” Through this training, the model learns patterns of word distribution, syntactic regularity, sentence length, lexical predictability, punctuation frequency, and other probabilistic features. When a user submits a text, the detector computes how closely that text resembles the statistical patterns of the AI-generated samples in its training set.

The result is therefore not a determination, but a probability estimate: a statement about how similar the submitted text is to what the model has learned to associate with machine-generated language. This process is no different in principle from how generative models themselves operate : both rely on pattern recognition across large textual datasets.

Crucially, this approach is inherently unstable. Human writing is highly diverse, and creative, edited, genre-specific, or non-native-speaker prose often overlaps strongly with the statistical profiles learned by these detectors. As a result, false positives are common and unavoidable. For this reason, in academic and forensic contexts, AI-detection tools are not regarded as reliable evidence of authorship; at best they provide weak heuristics, not proof.

Moreover, AI detection tools are commercial products. Their outputs are designed to appear authoritative in order to encourage continued use or paid upgrades, despite the underlying uncertainty of their predictions. You should therefore understand that submitting text to an AI detector is not performing an objective test, but rather feeding that text into yet another AI whose output is probabilistic and commercially motivated.

Conclusion: AI detectors do not “find” artificial intelligence within a text. They apply another AI model to estimate similarity to prior training data. Treating such estimates as definitive proof is a category error; one that risks misidentifying human authorship and undermining trust in creative communities.

AO3 Dataset

Source: https://www.reddit.com/r/AO3/comments/1k6ie6v/ao3s_data_was_scraped_for_ai_what_to_know/

In April 2025, Archive of Our Own (AO3) became the subject of an unauthorized scrape in which a registered user collected a large dataset of publicly available works from the Archive. This included fanfiction and artwork that was accessible without login. AO3 administrators were made aware of the incident and took action, including issuing a DMCA takedown request against a version of the dataset that had been uploaded to Hugging Face. That particular upload was made temporarily inaccessible, though copies of the dataset had also been posted elsewhere. Importantly, only publicly available works were scraped (works that were visible to anyone without account login). Archive-locked works (restricted to registered users) were not included in the dataset. AO3 staff and volunteers have long expressed concern about widespread data scraping, including how it can feed into large language model training, and have implemented measures like rate limiting and opting out of major crawls used for training datasets.

While some of the scraped datasets have since been taken down or removed from public distribution, the possibility that this material was used to train additional models remains entirely plausible.

This leads to an important conclusion: AI-generated writing can resemble fanfiction not because fanfiction is “becoming AI-like,” but because AI systems have been trained on fanfiction.

The resemblance runs in one direction only. It is therefore a fundamental error to assume that a fic “sounds like AI,” when in reality AI has learned to sound like us.

Showcasing with my own writing

Because I couldn’t help myself, I tested it. I deliberately copied and pasted my own writing into several different AI detectors to see what they would report. A piece of writing I know for a fact, has never been touched by AI.

The results were striking. The same passage of text produced wildly different outcomes depending on which detector I used. One flagged it as highly likely to be AI-generated, another judged it to be entirely human, and a third fell somewhere in between. None of these tools agreed, not on percentages, not on confidence, and not even on the basic classification. (See results below)

This is not a flaw of any single detector; it is a structural limitation of the entire approach. Each detector has been trained on different datasets and tuned to different models. Some are calibrated primarily against ChatGPT-style outputs, others against Gemini-like prose, others against mixtures of synthetic and human-edited text. As a result, they are not measuring an objective property of the writing itself, they are measuring similarity to whatever their own training data happens to look like.

In practice, this means that AI detection is not reproducible. A scientific test must yield consistent results when applied to the same input under the same conditions. These tools do not. They produce unstable, contradictory outputs because they are not detecting anything intrinsic to the text, they are guessing based on pattern resemblance.

What this demonstrates is that AI detectors do not provide evidence. They provide opinions: automated, statistically informed opinions, but opinions nonetheless. Treating those numbers as proof of AI use is therefore methodologically unsound, and in creative communities it becomes actively harmful, because it encourages suspicion based on tools that cannot reliably distinguish human authorship at all.

Quillbot:

ZeroGPT

GPTZero

Scribbr AI Detector

aidetector.com

mydetector.ai

copyleaks.com

reilaa.com

grammarly.com/ai-detector

Links

AO3 reddit thread of the scraping incident in April 2024: https://www.reddit.com/r/AO3/comments/1k6ie6v/ao3s_data_was_scraped_for_ai_what_to_know/