Language Log

"AI is doing job interviews now—but candidates say they’d rather risk staying unemployed than talk to another robot", Fortune 8/2/2025:

AI is replacing human hiring managers in job interviews—and candidates are pushing back. Despite being unemployed, professionals told Fortune they’re refusing to take calls with bots, calling it an “added indignity” and a red flag for company culture. Still, stretched-thin HR teams say it’s the only way to handle thousands of applicants.

Evaluations of these interview bots don't seem to have improved over the past year — see e.g. "Dystopian and dehumanising: that’s what it felt like when I was interviewed by an AI assistant", The Guardian 10/12/2026, and many similar reactions.

There are lots of other reasons for interacting with a chatbot, like it or not, and one of the things that often bothers people about those interactions is what's come to be called glazing, which means "to hype and compliment someone in a way that is over-the-top and cringey."

Over the past few weeks, Graham Dot Zip has created several amusing (fake?) interviews that feature glazing responses from job-interview chatbots, prompted by nonsensical inputs:

Mayur Naik has invented "MathDuels: A Self-Play Benchmark That Grows":

… in which models occupy dual roles: each authors math problems under adversarial prompting and solves problems authored by every other participant. A generation pipeline produces and filters the problems, and a psychometric model jointly estimates solver abilities and problem difficulties, from which author quality is derived. Experiments across 19 frontier models reveal that authoring and solving capabilities are partially decoupled, and that dual-role evaluation exposes capability separations invisible in single-role benchmarks. As newer models enter the arena, they produce problems that defeat previously dominant solvers, so the benchmark’s difficulty co-evolves with participant strength rather than saturating at a fixed ceiling. [Abstract taken from today's lecture.]

There's an evolving MathDuels leaderboard here.

Some of the comments on "Filthy comfortable" (10/9/2026) related to the issue of an adjective being used as an adverbial intensifier, which is a fairly common though always irregular development.

Wikipedia’s article on Benjamin Franklin starts this way:

Benjamin Franklin was an American polymath active as a physicist, philosopher, statesman and inventor.

The article goes on to list (some of) his other achievements, including his role as a writer, and as the founder of institutions such as the American Philosophical Society and what became the University of Pennsylvania. It omits many other things, including his contributions as an orthoepist and phonetician.

I spent yesterday afternoon at the Memorial Service for Victor Mair, where many of his relatives, friends, students, and colleagues told stories about Victor and his influence on their lives.

Because Grace Wu was not able to attend, she sent me a tribute to read on her behalf, which I've reprinted below.

A linguistically striking ad:

There's another one here, and the slogan alone also appears in text form.

Last Tuesday's post "Pneumatic, pneumonic, whatever" was based on a New York Post headline announcing that "Trump calls reported pneumatic plague outbreak ‘big problem for Russia’". It was implicit in the discussion that the headline writer was ignorant or careless or both.

But yesterday, Bob Ladd pointed out to me that Science magazine committed a similar error in the references to a paper dated 9/25/2026 (Morvai, Boglárka, Kinga G. Tóth, Marianna Boros, Dorottya S. Rácz, Ivaylo Iotchev, Kitti Szabó, and Attila Andics. "Neural evidence that dogs segment the speech they hear with a humanlike consonant bias." Science (2026).

Beth Simone Noveck, "How AI can be good for democracy", NYT 10/6/2026:

Governments can also use their enormous purchasing power to demand more from A.I. companies and to build more A.I. with them. In return for providing government funding, South Korea is securing contracts with multiple tech and telecom companies to give A.I. access to all of its 51 million citizens.

Giving AI access to everyone! That sounds like a good idea, right?

New information about the Memorial Service for Victor Mair, Saturday 10/10 from 2:00 to 5:00pm: For those who can't attend in person at  Swarthmore Friends Meeting (12 Whittier Place, Swarthmore, Pennsylvania), it will be available here through Zoom.

"Trump calls reported pneumatic plague outbreak ‘big problem for Russia’: ‘The microbes and the germs have gotten stronger’", NY Post 10/5/2026.

President Trump gave an ominous assessment of the suspected outbreak of pneumonic plague in Siberia Monday, calling it a “big problem for Russia” and adding cryptically that “the microbes and the germs have gotten stronger.”

A few weeks ago, I got an interesting email from Jean Charconnet, referencing a 9/10/2026 LLOG post, noting general issues about the logic of analogies and metaphors, and offering to submit a guest post on the topic.

Philadelphia seems to attract Google Maps weirdness. The last time I opened a map to verify some local directions, it popped up an image of the Grand Concourse ceiling of the 30th St. Amtrak station, identified as "Philidelphia Train ticket office":

Myriam Toua, "Donald Trump's advisers 'urge him to declare martial law' amid latest grim poll", The Mirror_us 9/26/2026:

[emphasis added] Martial law in the US would mean the temporary replacement of government and regular law enforcement by a solely military force.

Standard civil liberties, such as freedom of speech, assembly, or protection from detention without trial, are all suspended in the case of martial law being enacted.

The president has already tried to claim greater federal control over elections, which under the Constitution are administered primarily at the state and local levels of government.

In February, he said Republicans should nationalize boating in at least 15 places.