Tecof • September 15, 2026
Ethics in AI Content: Transparency, Attribution, Fact-Checking
In Brief
AI ethics is not an abstract debate but three concrete controls placed inside the content pipeline: disclosing what was produced with AI (transparency), showing whose work and which source it rests on (attribution), and checking every claim that goes live (fact-checking). What the three have in common is that they are solved by process, not by good intentions; "let's be careful" is not a control. As of 2026 the question is no longer "is it ethical to use AI" but "who approves which claim, and is that recorded".
Tuesday morning, 09.20. Eighty-four product descriptions for a cosmetics store were generated with AI and went live over the weekend. The first complaint on Monday quotes two sentences: "clears eczema marks" and "dermatologist approved". Neither is true of the product; both came from the model filling a gap it had no context for. Scanning all eighty-four takes four hours, and the legal exposure of those two sentences costs far more than four hours.
The problem was not the model's bad intent but that nowhere on the line did anyone ask "who approved this claim". Three weeks later the same team produced 240 descriptions with the same model and not one health claim shipped. The model had not changed; a banned-claim list, an automated scan and a single approval signature had been added.
Transparency: What to Disclose, to Whom, and Where
The transparency debate usually starts in the wrong place: "should we put an AI label on every page?" The right question is this: are you withholding something that would change the reader's decision? That a model drafted a product description does not change what a buyer decides. That an expert opinion, a customer review or a test result was generated changes it directly.
Content that needs disclosure and content that does not
The line is not about production method but about what the content claims. Who wrote "this shirt is 100% cotton" is irrelevant; whether it is true is not. A model writing "I wore this shirt for three months" is a fabricated testimonial, and labelling does not fix it — it should not be published at all.
| Content type | Disclosure needed? | Where | Why |
|---|---|---|---|
| Product and category copy | No | — | If the claim is true, the method does not affect the decision |
| Blog and guides | Optional | End of piece or byline | Transparency earns trust; it is not required |
| Expert opinion, case study | Yes | Inside the text | It carries a claim to authority |
| Customer review, testimonial | Do not publish | — | A generated testimonial is a misleading commercial practice |
| Support conversation | Yes | At the start of the chat | The person must know who they are talking to |
Saying you are a bot
If a chatbot answers instead of a person, saying so in the first line is both honest and in your interest: managing expectations reduces complaints. "Hello, I am the Tecof assistant; I can answer order and shipping questions and will hand you to the team if needed" makes the user's second message far more useful. The cost of hiding it is this: the user works it out eventually, and from that point they distrust the correct information too.
The regulatory side
Turkey has no specific rule mandating an "AI label", but three existing obligations touch this area directly. Under KVKK, automated decision-making must appear in your privacy notice and the model provider must be listed among your data processors. For the Advertising Board, the ban on misleading commercial practice ignores production method: a fabricated testimonial or an unsubstantiated claim is the same breach whether AI or a human wrote it. If you sell into Europe, the transparency articles of the EU AI rules ask for generated content to be marked and for bot interaction to be declared; for exporting stores that is now a compliance line item.
Attribution: Source, Copyright and Whose Work It Is
Attribution merges two separate questions: who owns the copyright in this content, and whose work does it rest on? The second is reputational rather than legal, and it costs more over time.
Copyright in model output
Raw model output with no human contribution tends to fall outside copyright protection; what you can defend as "our text" is the set of choices you made, the editing you did and the original data you added. The practical consequence: a competitor can produce something similar from the same prompt and you cannot object. The protectable asset is not the copy but the data under it — your own test results, your own sales breakdowns, your own customer records.
Citing sources and competitor content
Feeding competitor pages to a model and saying "write something like this" is a common and risky shortcut. Even when the output looks original, the structure, the ordering and the examples carry over — and so does the competitor's mistake. The healthy route is to use the source as data and attribute it plainly: "according to TÜİK's 2026 household ICT survey" is both verifiable and valuable. On the visibility-in-AI-search side too, the content that wins is the content whose sources are visible.
Images and brand assets
Teams that are careful with text get careless with images. Generated imagery has three boundaries: the face of a recognisable person, another brand's logo or packaging, and a protected character. Cleaning a background, adding a shadow or standardising framing on a product photo is fine; "make it look like that brand's campaign visual" is not. The in-house rule can stay simple: no real human face and no other brand's name in generated imagery.
Keeping human work visible
If a piece carries a "Tecof" byline, there is a person behind that byline. On our own line we use a split where the model produces the draft and a person produces the data and the final decision; we described how that was built in the piece on transformation on the agency side. The key point is that a byline is ownership: if something is wrong, it is clear who fixes it.
Fact-Checking: Keeping Fabrications Out of Production
The most expensive behaviour of a model is inventing what it does not know, and that behaviour surfaces exactly where you are weakest: missing product data. On a product whose attributes you never entered, the model fills the gap, because that is its job.
Three layers of review
One read-through is not enough, because a human eye stops seeing errors on the third pass over the same text. The three layers work like this. The first sits inside the prompt: the constraint "use only the attributes given; if something is missing, say it is missing". The second is automated: a scan against a list of banned words and claims — health claims, absolutes ("guarantees", "definitely cures"), competitor brand names, character-limit overruns. The third is human and looks only at what the second layer flagged; it does not reread every piece from scratch.
Which claim gets checked at which level
Checking every sentence with the same rigour is impossible; they have to be separated by risk.
| Claim type | Example | Risk | Check | Approver |
|---|---|---|---|---|
| Concrete product attribute | "100% cotton, 180 gsm" | Medium | Automated against product data | Catalog owner |
| Health or efficacy claim | "Clears eczema" | High | Banned list + mandatory human sign-off | Category manager |
| Figures and statistics | "40% of the market" | High | Primary source required | Content editor |
| Comparison | "Faster than rivals" | High | Removed unless measured | Content editor |
| Tone and phrasing | A promotional line | Low | Sampling, read 10% | Copywriter |
The table's real function is to show that the three high-risk rows are a small share of total copy. The way to manage review cost is not to read everything but to read the right five percent.
The source-required rule
A single rule closes most of the invented-statistic problem: a numerical claim does not ship unless it rests on a source you could link to. Asking the model to "add the source" is not enough, because it can invent the source too; you need a person who checks that the link opens. To make that check cheap at volume, it works better to ban numerical claims in the prompt outright and add the necessary ones by hand. We covered the logic of writing constraints into prompts in a separate piece.
Personal Data, Consent and Customer Content
This is the quietest but riskiest side of the ethics conversation: where customer data goes while you produce content.
Feeding reviews and messages to a model
Having reviews analysed is a legitimate use, but data minimisation applies: the review text and the order type are enough; name and order number are not needed. In practice the easiest route is a step that strips personal fields before the data reaches the model; that step is written once and runs every time. The same logic covers the whole of the review-analysis flow.
The model provider is a data processor
This sentence is unwritten in most companies and is the first thing asked in an audit. The provider you use must appear in your privacy notice and your processor inventory; your contract must state plainly whether your data is used for training; prompt and output logs must fall inside your retention policy. Enterprise plans usually carry a "not used for training" commitment and free plans usually do not, and that is precisely where the difference lies.
Consent in commercial messages
AI-personalised email and SMS sent without recorded consent is a breach regardless of how good the personalisation is. The key point: a prompt cannot check consent, a system can. The consent lookup has to sit in the send flow as a code step, and records without consent must drop out of the list automatically. We detailed how personalisation relates to consent in the email campaigns piece.
Writing the Policy in Thirty Days
Ethical rules that live only in conversations evaporate in the first busy week. The timeline below turns scattered instincts into a one-page policy someone can actually apply.
Days 1-7: take inventory
Which content types use AI, with which model, who runs it, who approves the output? No rules are written this week; the current state is. Most teams find two surprises when they list it out: a use nobody knew about, and an output flow nobody approves.
Days 8-14: write the banned list and the constraints
Build a banned-claim list per category — health claims in cosmetics, nutrition claims in food, performance guarantees in electronics. Put that list into the constraints section of your prompts and into the automated scan at the same time. This week's output is three files: the banned list, the prompt constraints, the scan rules.
Days 15-21: set the approval chain
Name a single approver for each content type and write it into the system. Content two people share responsibility for is content nobody is responsible for. For high-risk claims the approval step must be unskippable; low-risk copy passes on sampling.
Days 22-30: measure and version
Pick fifty pieces that went live last month and count how many needed a correction. That rate is the measure of your policy, and it gets remeasured every quarter. Keep the policy with a version number so it is visible who changed what. An ownerless policy loses touch with reality within three months.
Here is the job for tomorrow morning: open ten pieces you produced with AI and published in the last month, and underline every sentence where you cannot answer "how do we know this is true?". That list is the first draft of your banned-claim list. On a setup with AI built into the platform most of these controls live inside the flow, and the permissions you open to agents stay inside the same boundaries.
Frequently Asked Questions
Am I obliged to state that I used AI?
There is no general obligation in Turkey. The obligation comes from the type of content: it is needed where the content carries expert opinion, testimony or a claim to authority; it is not needed in accurate product copy. If you sell into Europe, check the EU transparency obligations separately.
Does AI-generated content hurt SEO?
The production method is not penalised on its own; worthless content is. Hundreds of pages cut from the same template with no original data will hurt. Content that adds your own sales data, your own test results and real examples holds up regardless of who wrote it.
Can I use AI to summarise customer reviews on the page?
Yes, as long as it is clearly a summary. A block headed "the three things customers mention most" is legitimate. What is not legitimate is turning that summary into something that reads as one named customer's own words; that is a fabricated testimonial.
How do I stop the model inventing statistics?
The most effective method is banning numerical claims in the prompt and adding the necessary figures by hand. "Cite your source" is insufficient, because the source can be invented too. Make it a rule that every figure rests on a link that opens before publication.
Does admitting the chatbot is a bot reduce sales?
Field experience points the other way: with expectations set correctly, users write clearer questions and resolution rates rise. What does cause a drop is a bot presenting itself as human and then failing to help; the user is disappointed twice.
Can I use people in AI-generated imagery?
Do not use a face resembling a recognisable real person. A wholly invented face is technically possible but in fashion and cosmetics it risks distorting how the product actually looks, and your return rate will bill you for that risk. Background and lighting work on product photos is the safe zone.
Who should own this policy?
The team producing the content, not legal or technology. Legal draws the boundary, technology supplies the tooling, but the daily decision belongs to whoever publishes. A policy whose owner sits outside the production team does not get applied.
Isn't this much control excessive for a small team?
On a three-person team the policy is one page, the banned list is one table and the approval chain is one name. What is excessive is not the control but making it complicated on paper. The test: can a new joiner read these documents and start working?
Does telling customers we use AI reduce trust?
How you say it decides. "Our content is produced by AI" reduces trust; "we speed up drafts with AI and our team verifies every product detail" increases it. What makes customers uncomfortable is not automation but the absence of control.