Can a teacher tell if a student used it?

Not reliably, with a detector. A detector doesn’t find a hidden mark in the text. It makes a statistical guess. Most detectors, one research team notes, lean on how predictable the words are, a measure called perplexity: how surprised a model is by each next word.¹ Chapter 3 showed Shannon measuring the predictability of English in 1951.² Plain, predictable human writing can look machine-made by that measure.

OpenAI released its own detector in January 2023 and called it “not fully reliable.” On its test set it caught 26 percent of AI-written text and wrongly flagged human writing 9 percent of the time. In July 2023 OpenAI withdrew it “due to its low rate of accuracy.”³ A 2023 study tested seven detectors on 91 essays by non-native English speakers for the TOEFL exam. The detectors flagged more than half as AI-written, an average false-positive rate of 61.3 percent, while scoring U.S. eighth graders’ essays almost perfectly.¹ An international team that tested fourteen tools found them “neither accurate nor reliable.”⁴

“A detector score proves cheating.” Even OpenAI said its tool “should not be used as a primary decision-making tool.”³ The researchers behind both studies advise against using detectors to judge students.¹⁴ Those studies tested 2023 tools; newer ones may do better or worse, so ask for independent test results, including on people writing in a second language.

Students do use these tools. In a Pew Research Center survey in fall 2025, 54 percent of U.S. teens said they had used chatbots to get help with schoolwork. One in ten said they did all or most of their schoolwork with a chatbot’s help.⁵ Chapter 12’s habit applies to teachers too: look for evidence you can check, such as drafts and a conversation with the student. Chapter 13 covers saying plainly how AI helped.

Could a voice on the phone be fake?

Yes. A familiar voice is no longer proof of who is calling. “A scammer could use AI to clone the voice of your loved one,” the Federal Trade Commission warned in March 2023. All it takes, the FTC said, is a short clip of the person’s voice, perhaps from something posted online, and a voice-cloning program.⁶ The FBI warned in December 2024 that criminals generate short audio clips in a loved one’s voice to pose as a relative in a crisis and ask for money right away.⁷

The defense is old-fashioned. “Don’t trust the voice,” the FTC says. Call the person back, using “a phone number you know is theirs.” A request to wire money, send cryptocurrency or buy gift cards could be a sign of a scam.⁶ The FBI suggests a secret word or phrase that your family agrees on ahead of time.⁷ At work, the same trick aims at money. The Treasury’s financial-crimes unit reported in November 2024 that financial institutions were filing more reports of suspected deepfakes in fraud, often fake identity documents.⁸ The rule is the same as at home: confirm through a channel you already trust, such as the number in the company directory.

The crimes aren’t new, and old laws are being applied to them. In February 2024 the Federal Communications Commission ruled that a 1991 law limiting robocalls with “artificial or prerecorded” voices covers AI-generated voices too.⁹ In September 2024 it imposed a $6 million penalty, under a law against faking caller ID, on Steve Kramer for robocalls before New Hampshire’s 2024 primary that carried a deepfake of President Biden’s voice. One commissioner dissented.¹⁰ Both are agency actions, not court rulings.

“You can always spot a fake by its glitches.” Don’t count on it. The Treasury alert, drawing on the Department of Homeland Security, notes that deepfakes “do not need to be particularly advanced or believable to be effective.”⁸ Chapter 12’s rule applies. Check the claim through a source you trust, not through the message itself.