"Generative AI" != speech-to-text, diarization etc.
« previous post | next post »
In a July 23 Memorandum Decision from the Court of Appeals of Indiana (noted by Robert Freund and 404 Media), Judge Jeffrey L. Marchal complains that
The Transcript contains various types of errors. There are numerous typos that change the meaning of the testimony, question, or objection. See, e.g., Tr. Vol. II at137:18, 144:10, 147:10; Tr. Vol. III at 6:13. In some instances, witnesses’ and trial attorneys’ names are reported incorrectly. Tr. Vol. II at 220:5; Tr. Vol. III at 142:15–20, 143:15, 162:4–5. At one point in the Transcript, a motion, presumably made by the State, is attributed to the trial court. Tr. Vol. II at 107–08. At another point, an objection, presumably made by Williams, is attributed to the Bailiff. Tr. Vol. II at 177:15. At yet another point, the State’s closing argument is attributed to the trial court. Tr. Vol. III at 228:1. […]
Based upon the types of errors reviewed, it appears that generative artificial intelligence may have assisted with the preparation of this transcript. While AI can improve efficiency and be a productive tool for many professionals, it is incumbent upon those using such systems to proofread and ensure the accuracy of the generated product.
Almost certainly, the cited problems are the result of errors committed by a speech-to-text system, which the court reporter used to create a draft transcript and failed to correct.
The language models used in speech recognition, and the classification methods used in diarization ("who said what"), are technically similar to the generative models that modern chatbots use to create text, but they're not really being used generatively in those applications.
The term artificial intelligence has been long since been extended to mean little more than "complex computer program" — and apparently "generative artificial intelligence" is going down the same path. It's rather like using tilling, sowing, fertilizing, weeding, watering, and harvesting interchangeably to cover anything having to do with farming…
JPL said,
July 25, 2026 @ 5:02 pm
Strangely, the senses of Chomsky's uses of the term 'generative' have changed over time between his allusions to Humboldt and the minimalist program. The latter seems to bear no traces of any property that could be called "generative" in his original sense. In the public mind it looks like it's the term 'generative' that is going down that path to semantic vaporization. I want to look more closely at the changes in Chomsky's conception of what the property of "generativity" is. It's a centrally important idea in understanding the logical structure of language, as well as in mathematics, but the locus of its operation in the phenomenon of language has not been correctly identified. It's not equivalent to "recursivity", for example. How do linguists today understand the term 'generative'? What property of the phenomenon of language (usage and system) does this term properly describe?
(" … but they're not really being used generatively in those applications." I would suggest that when talking about the property of generativity belonging to the "models" that chatbots use to create text, it would probably be better to use a term other than 'models' to identify whatever it is that (supposedly) has this property. It's a logical relation that needs to be specified.)
Chester Draws said,
July 26, 2026 @ 5:38 pm
The term artificial intelligence has been long since been extended to mean little more than "complex computer program"
That's a bit rough. No-one really thought Deep Blue was intelligent. No-one thinks those massive meteorological programs that predict the weather are intelligent.
I would argue that "AI" is generally referring to any program that solves different problems in multiple contexts. I would guess that most people would expect them to be a bit probabilistic too.
Still not intelligence, of course, but at least mimicking it.
HTI said,
July 27, 2026 @ 3:42 am
I don’t see any motivation for supposing that the meaning of “AI” might have anything to do with the software as such. It’s pretty clearly a purely sociological term.
If I had to gloss it off the top of my head, it’d mean something like “whatever software architecture is both sufficiently new and about which we can plausibly build a narrative tying it to the sci-fi trope of computer programs that can think”.
The facts that (1.) a computer program isn’t even the kind of thing that can think, and that (2.) technological trends are everchanging in search of more military funding or (more recently) VC funding, probably jointly contribute to the term “AI” necessarily being an ever-moving and — from the software point of view — terribly ill-defined target.
Kakurady said,
July 28, 2026 @ 8:04 am
While speech recognition and generative AI are different applications, nowadays several popular speech recognition models are based on the same "transformer" technique as the large language models that power generative AI, and are released by the same organizations that produce those large language models.
In some cases, a large language model is used as part of speech recognition application, and there are also "multimodal" generative models that can be used with audio or text input, instead of specialized speech recognition algorithms.
So depending on the underlying technology, it may not be inaccurate to say generative AI caused the transcription error. It is also possible that the court reporter used an AI assistant to revise the transcript. Of course, these are mere speculations without knowing the exact software used to produce the transcript.
JimG said,
August 2, 2026 @ 12:56 pm
There's a fairly straightforward differentiation of generative and speech recognition AI, apparently by Quora's AI, at
https://qr.ae/pFb8B1