Does evidence have to prove a hypothesis?

No. It can raise or lower a probability while leaving uncertainty.

Is a noisy sensor reading still evidence?

Yes, if the inference model accounts for how readings relate to the underlying state.

Can the same evidence lead to different posteriors?

Yes. Different priors or likelihood assumptions can produce different conditional results, so those assumptions should be inspectable.

AI Glossary / Graphs and symbolic AI / Intermediate

Evidence

Definition

Evidence, in probabilistic AI, is the information on which an inference is conditioned. It commonly takes the form of observed values of variables, such as a sensor reading or a recorded event. A model combines this evidence with its prior assumptions to update beliefs about variables that remain unknown.

What it means

The word has a broad everyday meaning, but its technical role is specific. In a query such as P(H given E = e), H is a hypothesis or unknown variable, and E = e identifies the observation being supplied. The resulting probability is conditional on that observation and the model.
Calling something evidence does not certify that it is reliable or decisive. A recorded signal may be noisy. The model can represent that noise by distinguishing the observed reading from the underlying state. This avoids treating a fallible measurement as direct knowledge of the hidden condition.

How it works

Identify the observations and the variables they constrain. Combine their likelihood under possible hypotheses with the prior distribution, then normalize to obtain posterior probabilities. For a discrete hypothesis H and observed event e with positive probability, Bayes' rule gives P(H given e) = P(e given H)P(H) / P(e).
With several observations, their dependence matters. Repeated copies of one report are not automatically independent evidence. A Bayesian network can make the assumed relationships explicit. Some systems also accept uncertain or likelihood-based evidence, but its interpretation must be specified rather than silently replacing a definite observation with a guessed value.

Three practical examples

A monitoring prototype updates its belief about a machine state after receiving a noisy indicator reading.
A document-classification exercise conditions on observed words while keeping the document's category unknown.
A tracking model combines a sequence of position measurements with assumptions about how the hidden position changes over time.

Think of it this way

Think of a detective narrowing possibilities after receiving a clue. The clue can favor some explanations without proving one. The analogy is limited because probabilistic inference requires explicit likelihoods and dependence assumptions; an AI model does not automatically possess the contextual judgment needed to assess every real-world source.

Worked example

Consider a fictional machine indicator. Let H mean the machine is in a special operating mode. Before reading the indicator, P(H) = 0.2. Suppose a blue signal occurs with probability 0.9 in that mode and 0.1 otherwise.
The probability of a blue signal is 0.9 × 0.2 + 0.1 × 0.8 = 0.26. After observing blue, the posterior probability of H is 0.18 / 0.26, approximately 0.6923. The signal changes the belief from 20 percent to about 69.2 percent, but does not make the mode certain.
Imagine 1,000 hypothetical cases matching this model exactly in expected counts. About 180 would be special-mode cases with blue signals, and 80 would be other-mode cases with blue signals. Among the 260 blue signals, the expected special-mode proportion is 180/260. These are illustrative model expectations, not observed performance figures or operational guidance.

Illustrative case study

In a fictional workshop, the Elm team receives two dashboard alerts and initially treats them as two independent confirmations of a hidden condition. An investigation of the simulator shows that both alerts were generated from the same underlying sensor reading.
The team revises the observation model so the duplicated displays do not count as independent measurements. It documents which event was actually observed and which values were copied downstream. The case illustrates evidence provenance and dependence, without claiming that a particular posterior estimate is correct for a real machine.

Common misconceptions and mistakes

Do not confuse P(evidence given hypothesis) with P(hypothesis given evidence). The worked example's 0.9 sensor likelihood differs from its approximately 0.6923 posterior. Prior probabilities and alternative explanations matter.
An absence of a recorded event is not automatically evidence that the event did not occur; the recording process may be incomplete. Likewise, more observations can worsen an inference if the model misrepresents their dependence or reliability. A posterior is a conditional result within the model, not a guarantee that the observations or assumptions are accurate.

Business use

The concept helps teams explain why an AI estimate changed. A useful explanation identifies the new observation, the hypotheses compared, and the assumptions governing the update. It should separate observed inputs from inferred states.
For operational review, preserve provenance and timestamps where they affect interpretation. Check whether the system is repeatedly counting one source, conditioning on outdated information, or treating missing values as definite negatives. These questions are about the meaning of the input to inference, not simply the amount of data available.

Check your understanding

The blue indicator is 90 percent likely when H is true. Why is P(H given blue) not automatically 90 percent?

Answer: Blue can also occur when H is false, and H initially has probability 0.2. Bayes' rule includes those alternatives and the prior, producing about 69.2 percent in this example.

Related terms

Learn next

Learn next about likelihood and posterior probability. Recalculate the blue-signal example with prior P(H) = 0.5 while leaving both sensor likelihoods unchanged. Explain which input changed and why the resulting posterior differs, then consider what extra assumption would be needed to combine two sensor readings.

Sources

Russell and Norvig, original probability slides, conditional probability and evidence: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter13.pdf Kanazawa, Koller, and Russell, original paper on observations in dynamic probabilistic networks: https://arxiv.org/pdf/1302.4965

Previous: Event tree analysis Next: Evidence reversal