AI tone analyzer without a taxonomy

The output is a written account of how a text is likely to land on a reader: which passages set the register, where the register diverges from the situation, and what a specific phrase communicates beyond its literal content. It is addressed to someone about to revise the text.

The route is server-side. Text is transmitted to our AI server, input is accepted between 1 and 12,000 characters, the instruction field holds 300 characters, and the allowance is 5 free runs per rolling 24-hour window with no account.

The established tools here are classifiers. One claims more than 25 distinct tones, naming joy, confusion, gratitude and neutral among them, and assigns a tone to the document and separately to every sentence. Another sorts text into 12 categories organised as professional, casual and emotional families. Classification is cheap to consume, comparable across documents and trivially chartable, which is a genuine advantage.

It also has a ceiling. A sentence-level label of neutral does not identify that the third line reads as a reprimand because it opens by restating what the recipient got wrong. The label is correct and useless simultaneously. This route produces no taxonomy at all, which forfeits comparability and buys specificity: the passage is named and the mechanism is explained. Neither model dominates, and a page that sells one while implying the other misleads its reader.

Submit the complete message. Greeting and sign-off carry disproportionate tonal weight and are frequently the passages responsible.

Step 1

Paste the full text, between 1 and 12,000 characters.

Step 2

Specify the recipient and their current disposition in the instruction field, then execute. The same sentence lands differently on a neutral reader and an aggrieved one.

Step 3

Revise the identified passages. Remaining runs are shown on the interface.

None of these is a grammatical fault, which is precisely why correctness-focused tools pass the text through unflagged.

Abrupt length contrast

A short sentence immediately after a long paragraph reads as patience expiring. The signal is the delta, not the length.

Restated error as opener

"As I mentioned" and its variants function as corrections irrespective of the writer's intent, because they foreground the recipient's failure to absorb prior information.

Agentless passive around fault

"The file was not attached" removes the agent syntactically while leaving it unambiguous, which reads as an accusation the writer declined to make openly.

Formality substituting for courtesy

Extended constructions defer the request. Courtesy is a clear ask plus acknowledgement; length is not a substitute for either.

Register held constant across content

An upbeat close attached to an outage notice reads as a failure to understand the situation rather than as positivity.

No tone label and no tone list, because no taxonomy exists in this codebase. No per-sentence tone assignment, since the route performs no segmentation. No sentiment value, no emotion score, no confidence figure, because nothing in this path computes a number.

No browser extension, no Gmail or Outlook integration and no API, all of which competing tools in this result set ship. No rewriting into a selected tone, which is a separate route here rather than a bundled feature. And unlike at least one competitor advertising unlimited free analysis, this route is metered at 5 runs per rolling 24 hours and the text is transmitted, so neither unlimited nor private would be a truthful description of it. A different product competing for the same keyword analyses acoustic properties of recorded speech, including pitch contour and speaking rate; this route has no audio path of any kind.

A model optimised for broad acceptability settles into the register that minimises objection: even sentence length, hedged assertions, a positive close independent of subject matter. It is the register of institutional announcement, and it is produced for a product launch and a service failure alike.

The resulting text reads as insincere rather than impolite. Tone fails to track content, and readers register the divergence between the news and the voice delivering it well before they can articulate what is wrong.

No. Tone is an observable property of text, not evidence about its origin. Institutional human writing occupies the same flat register routinely, and edited model output frequently does not. The classifier is instructed never to assert certainty about authorship.

We make no prediction about any third-party detector's output, do not tune against detector scoring and do not test for it. The report is a revision instrument.

Writers immediately before sending a message whose reception is consequential: a complaint, an escalation, a refusal, a correction to someone senior. The value concentrates at exactly the moment the stakes rise.

Support leads validating a template that will reach a large recipient set, where a single misjudged construction is multiplied rather than isolated, and communications teams checking that an announcement carrying difficult news does not read as either cold or inappropriately bright.