Skip to content
Interface language: English
Audit / Signal

AI returns to its own suggestions: how do you know the data is new?

Trace an AI label through instructions, knowledge and evaluation without mistaking the tool’s old opinion for independent evidence.

Article language: English

An illustration of a gold-coloured lantern lighting stone steps and cracks in a wall.
In this article

An assistant labels a message as an ordinary installation question. The label goes into the CRM, then into an export of "resolved cases", and finally into a set of examples for the same assistant. A month later the tool again recognises a similar message as an installation question. The team treats that as confirmation of quality. In reality, the AI has mostly encountered its own earlier opinion, now wearing a shirt labelled "data".

This is a fictional cycle at Rytm, a company selling modular fixtures. The original message concerned a missing item from the kit, but nobody checked it. We do not need to assume the model was retrained. Its label can return as an example in instructions, an entry in a knowledge base or a criterion for evaluating a new output. The information changes status: a suggestion starts to pass as an independent observation.

Here, feedback means a cycle in which an output affects material used in later actions. Returning an output is not necessarily an error. Suggestions, corrections and accepted examples can be used deliberately. The problem begins when provenance or the meaning of an assessment disappears along the way. Another column called "quality" will not fix that if it is filled with the same unchecked label.

Choose one category and follow it from the message to its next use. Record who assigned the first label, what closing the case means and the rule that places the record in the export. Starting with a discussion of all the company's AI is too broad. One message lets you see where the tool received its own answer as supposedly new evidence.

What exactly was approved?

In Rytm's sample, distinguish an AI suggestion, acceptance for the current case, a human correction, the case resolution and an assessment of suitability for future examples. These need not be separate screens, but they should remain recognisable events. Clicking "close" does not explain whether the case owner checked the category or merely ended the conversation.

Establish what the person actually confirms. They may approve sending an answer without judging the label. They may change a category for team organisation without assessing the meaning of the message. Such a decision can be useful operationally without necessarily becoming a good classification example. Testing AI message classification (in Polish) needs a clear definition of the category being evaluated.

For every example retain at least a link to the input, the original suggestion, the final resolution and the reason for a change. Do not overwrite the first label in a way that removes evidence of the tool's error. A later report should distinguish an accurate suggestion from a case repaired by the owner. The final category may be correct in both, but the assistant's work needs different assessment.

Prepare three controlled records: a suggestion left unassessed, a confirmed category and a category corrected after the customer clarified the case. Run a test export using the actual rules. Check which records were selected and how their provenance is described. If all receive a "verified" status, the problem is already in data preparation before the next model run begins.

Also inspect a record reopened later. Initially the customer accepted instructions, but later confirmed a missing part. If the export kept only the first closure, the dataset does not reflect the final resolution. Establish the point in time covered by the data version and how later corrections are carried across. There is no need to pretend today's state was known from the start.

Consider a correct AI label too. If the owner read the message, checked the conditions and confirmed the category, the example may be useful. Keep both events: the tool's suggestion and the subsequent assessment. Removing every trace of AI assistance from the material does not replace checking its basis. What matters is where the resolution came from and what the person actually assessed.

A loop can operate without retraining

In Hidden Technical Debt in Machine Learning Systems, the authors describe direct influence by a model on future training-data selection and harder-to-notice dependencies between systems. This is a 2015 analysis of mechanisms in maintaining machine-learning systems, not evidence that every company automation has the same problem.

At Rytm, the simplest return path goes through instructions. Each week an owner adds "the most common correctly resolved examples" from an export. If selection relies on unassessed AI labels, the tool receives them as patterns. Changing the instruction version does not then mean independently confirmed knowledge has arrived. During review, open a specific example and check its source.

The second route runs through searching past cases. The assistant finds its own old summary saved without provenance. Several similar summaries can sound like a consistent customer history even though all were produced by the same mistaken rule. Check not only whether retrieved text fits the question, but whether it describes a separate event and a confirmed resolution.

The third route concerns quality assessment. New labels are compared with old ones, and their agreement is reported as accuracy. If the older labels were also model suggestions, the result primarily measures agreement with previous outputs. It cannot by itself establish correct categories. Call the measure agreement until an independent basis has been checked.

Aeneia conceptual photography

Case selection works differently again. If owners open only the "missing part" category, messages incorrectly labelled as installation questions may never be assessed. The later correction register shows only what someone had an opportunity to see. In a limited sample, examine cases outside the selected queue too. Do not treat the absence of correction as proof that a person confirmed the result.

A similar problem arises when the suggestion itself influences the owner's later record. In a trial, compare an assessment made with the label visible with an assessment of the material without that suggestion, using the same category definitions. A discrepancy calls for discussion of the decision's basis; it does not immediately measure the size of the effect across the company. "A person clicked" does not explain what they read or how they understood the choice.

At this point examine drift in incoming data (in Polish). A change in category proportions may reflect new customer needs, but also a new routing method. Compare the raw input, the tool's output and material available to owners. Otherwise it is easy to tell a market-change story based on a change in visibility inside your own system.

The study AI models collapse when trained on recursively generated data analyses a different, specific mechanism: successive generations of models trained on data produced by earlier models. It is not an automatic diagnosis of "model collapse" in a CRM. In our example, an error in information status can be shown without proving a change in model parameters or long-term degradation of its abilities.

If the company really prepares data for further training, expand the review to provenance and selection, as in preparing a fine-tuning dataset (in Polish). Do not mix that decision with a weekly instruction update. Each route needs its own record of which data returned and in what role it will be used.

Check intermediaries too. A consolidated spreadsheet might combine CRM records, survey responses and an automated report, then remove the source column. The person writing instructions sees a single list. Choose an entry from each route and trace it back to the original material. Two identical texts in different rows may be copies of one output, not two independent cases.

Stop the error returning while keeping useful corrections

A first fix may be an acceptance condition for the dataset. Closing a case alone is not enough; the category needs assessment under a defined rule. Assign an owner for that decision and identify cases that go for clarification. Do not transfer the whole burden with a vague instruction to "check everything" if the screen does not show what the owner is meant to confirm.

Test the correction end to end. In controlled material, change the installation-question label to missing part and add a reason. Then check the new export, retrieved material and sample used for assessment. If the old suggestion still returns in one place, the correction visible in the CRM did not close the loop. Identify the specific copy needing an update in the report.

If another team uses that copy, assign its owner and a way to pass on the correction. The person correcting the CRM may not have access to the training material. Do not mark the task complete simply because a notification was sent. In the sample, check the confirmed state of the destination document and identify its version. Handing over responsibility is a step in the flow, not evidence that the correction was made.

Provenance should survive movement between tools. A spreadsheet row, an instruction excerpt and an export file may have different formats, but should still allow the example's source to be recognised. Check this at an actual transition between tools. A folder called "checked" does not replace information about assessment of a particular record, especially when someone copies only selected columns.

Include instruction updates in acceptance too. Part of the team may still use an old version while the new one contains the correction. Record which version the examined run used and whether the earlier example remained available by another route. Without that distinction, two different outcomes can easily be described as model randomness when the tools simply received different material.

Mark synthetic data (in Polish) separately. It may be useful in controlled trials but is not a new customer-service event. If it supplements a particular rare case, record that purpose and the expected result. Its role should not change after export into "this is how customers write to the company".

Do not use the material being repaired as the independent test of that same fix. Rules for protecting a test set from leakage (in Polish) help keep separate assessment material. Rytm can prepare new, independently resolved messages about missing parts that were not used to shape the instructions. The trial should show behaviour on another input.

Attach a comparison of two things to the decision: correctness of new outputs and work needed to check them. Greater agreement with the company dataset may reflect genuine improvement or a dataset made more like old AI suggestions. Keep a stable sample with independent assessment and examine fresh cases separately. Present both results separately, including case counts and detected errors.

In the acceptance report, record the return path found, the rule for accepting examples and evidence that a correction travelled through the flow. The possible scope of a review of a sample from a chosen process can be discussed with Aeneia Audit. At Rytm the improvement is concrete: the assistant still uses earlier cases, but does not present its own old, unchecked label as fresh confirmation from a customer.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear