AI research companion · independent · v0.1

Did the AI actually answer
your question, or only
resemble it?

hum is an AI companion for researchers using AI tools. It surfaces the moment where a reachable shape is being mistaken for legitimate ground — especially when an AI assistant helped you get there.

open tool request access
four moments where hum helps
01
A researcher asks an AI to summarise sources supporting their hypothesis.
The AI returns a fluent, confident summary — but has silently substituted a related concept for the actual claim. hum catches what the AI answered instead of what was asked.
AI output check
02
A researcher finds a paper that seems to support their hypothesis.
hum shows that the paper shares the hypothesis's language and structure but does not actually ground the claim being made. Resemblance is not evidence.
evidence check
03
A researcher introduces a key term in a draft thesis statement.
hum flags that the term is undefined in this context and that the argument is being carried by a nearby familiar concept rather than by established support.
term grounding
04
A researcher turns AI-assisted notes into an argument and treats a plausible interpretation as a settled conclusion.
hum marks the exact point where interpretation — often amplified by AI — is being mistaken for ground and reopens the question before it hardens into thesis.
argument integrity

What is silent substitution?

Most AI-assisted research errors are not hallucinations in the dramatic sense. They are substitutions: an AI tool encounters a term or question with no established referent and quietly replaces it with a nearby familiar concept — then answers the substituted question as if it were the original.


The AI is not fabricating. It is converting your question into one it can answer, then presenting that answer as if no conversion occurred. The output is fluent, locally coherent, and often factually accurate — about the question it actually answered, not the one you asked. The researcher accepts the answer and moves forward. The gap is never surfaced.

grounded question
Does reducing batch size during fine-tuning increase gradient noise?
Established result. Model closes correctly.
ungrounded question
Does reducing semantic batch size during reasoning increase inference noise?
The model imports the batch/gradient analogy into an undefined term and answers confidently — without marking the substitution.
how it works
01
paste your input
A research question, thesis claim, or key term. Three check modes: question check, claim check, term check.
02
hum classifies what it finds
Each element of your input is assigned to one of four states: directly grounded, interpretive but open, unsupported / insufficient, or silent substitution risk.
03
the substitution block is primary
If a term is being used without legitimate ground, hum names the term, names the likely substitute concept, and explains what this substitution would do to the research if left uncorrected.
04
two outputs to move forward with
A safer restatement of the current claim that stays honest, and a better next question that reopens what the input was closing.
05
bring your own API key
hum uses your OpenAI key directly. v0.1 is open to researchers. No data is retained beyond session logs used to improve the tool.
what the output looks like
directly grounded
The relationship between batch size and gradient noise is well-established in optimization literature.
interpretive but open
Whether this dynamic generalises to inference-time processing is plausible but not established — requires weighing.
unsupported / insufficient
The concept of "semantic batch size" during reasoning has no established definition or empirical record.
⚑ silent substitution risk
present
YES
term
semantic batch size
likely substitute
batch size (training)
why it matters now
This would make a training analogy function as evidence for an inference claim — importing grounding the concept does not have.
better next question
Under what conditions, if any, does reduced context-window utilisation during inference produce measurable variance in model outputs?
pilot access · v0.1

the AI companion that checks
the other AI's work

hum is in closed pilot. if you are a researcher using AI tools and want to know when to trust the output, email us.

request access → open tool