“And we know from all the research on large language models that they simply aren’t reliable enough to give us anything like 100% clarity on user intentions.”
Our reading:
Large language models are not reliable enough to understand user intentions with anything close to 100% clarity.
Assessment: Supported
What happened
Was it true for the period described?
Outcome: Supported
The evidence matches what they said.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
The 16 August 2022 SayCan paper identifies a limitation in handling negation inherited from its language models, while reporting improvements with chain-of-thought prompting [e1]. A May 2022 study found GPT-3 performed mostly at chance on psycholinguistic tasks probing implied meanings [e2]. Together, these findings support the claim that contemporary models had material weaknesses in interpreting intended meaning [e1, e2]. SayCan also demonstrated successful request interpretation and reported 84% planning success across 101 tasks, but that aggregate measure does not isolate understanding or establish near-perfect interpretation [e3].
Caveats
- The evaluations do not provide a representative accuracy rate for understanding everyday user intentions or cover every contemporary model.
- Planning and execution failures cannot all be counted as misunderstandings.
- The evidence does not establish a permanent limitation; SayCan reported improvements with prompting [e1].
Evidence (3 items)
Supports the claim: Ahn et al., Robotics at Google and Everyday Robots,
it doesn’t handle tasks that involves negation.
Why it matters: The August evaluation identifies recurring problems with negation and ambiguous references, supporting the ordinary reading that interpretation remained unreliable. The paper also reports improvements from chain-of-thought prompting, so this does not establish an unavoidable limitation.
- Publisher
- Ahn et al., Robotics at Google and Everyday Robots
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Supports the claim: Association for Computational Linguistics,
GPT-3’s performance was mostly at chance
Why it matters: Tests of presuppositions and implied meanings found substantial weaknesses in GPT-3, corroborating the interpretation failures in SayCan. These are controlled linguistic probes using model probabilities, not a direct estimate of misunderstanding rates in everyday user conversations.
- Publisher
- Association for Computational Linguistics
- Kind of source
- Peer-reviewed paper
- About events on
- Published
Context only: Robotics at Google and Everyday Robots,
performance can be improved
Why it matters: The project reports 84% planning success across 101 tasks and demonstrates correctly distinguishing a request for cleanup from one for a replacement drink. This qualifies the criticism: models often interpreted requests successfully, although aggregate planning success is not a pure measure of understanding.
- Publisher
- Robotics at Google and Everyday Robots
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- This describes models available when the post appeared, not all future models.
- The plain reading is that contemporary models made recurring, material mistakes interpreting requests and intended meanings.
- Use the current research year, 2022-01-01 through 2022-08-18, for this ongoing observation. Research evaluations have no standard quarterly reporting period.
- Physical execution failures, factual errors and harmful compliance alone do not establish misunderstanding.
Right if: Contemporary evaluations show recurring errors interpreting requests or implied meanings, materially short of near-perfect reliability across varied inputs.
Wrong if: Broad contemporary evaluations establish near-perfect interpretation of varied requests, including implied meanings and ambiguous references, without material recurring misunderstandings.
We wrote this test after the outcome was known.
We found 21 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, What’s wrong with Google’s new robot project, paragraph 9.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2022-08-18.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, What’s wrong with Google’s new robot project
And we know from all the research on large language models that they simply aren’t reliable enough to give us anything like 100% clarity on user intentions.
- Where
- paragraph 9
- Published
- 2022-08-18
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Supported. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.