What the evidence on AI personalisation and trust does and does not show
The research base on consumer trust in AI personalisation is thinner than the brief assumed. The regulatory and survey evidence that does exist points to disclosure and human oversight.
No peer-reviewed study of consumer trust in AI-driven personalisation appears in the sources we reviewed, and that gap is itself the finding. The nearest hard figures come from regulation and from national surveys. Pew Research Center reports that 72% of US adults say it is extremely or very important that providers tell them when AI is used in their healthcare. That figure concerns disclosure in health, not product recommendations, so it cannot carry the weight a board might want to place on it.
- US adults who say AI-use disclosure in healthcare is extremely or very important (Pew)
- 72%
- US adults who say AI will worsen the patient-doctor relationship (Pew)
- 47%
- Practices the EU AI Act bans, per the European Commission
- 9
What the data shows
The EU AI Act is the one binding framework in our sources. The European Commission describes Regulation (EU) 2024/1689 as the first comprehensive legal framework for AI, with the aim of fostering trustworthy AI in Europe. According to the Commission, the Act bans nine practices. Prohibitions one to eight took effect in February 2025, and prohibition nine follows in December 2026, added through the AI Omnibus. The Commission's page lists the banned practices. A lawyer should check whether any personalisation feature sits near one of them.
The transparency duties matter more for disclosure. The Commission says chatbot users must be told they are dealing with a machine, so that they can decide with information. Generative AI providers must make AI-generated content identifiable. Deepfakes and certain public-interest AI text must be clearly and visibly labelled. The Commission sets August 2026 as the date the transparency rules take effect, so a company that has not yet mapped its chatbot and content tools to these duties is working late.
Most EU AI systems carry no AI Act obligations at all. The Commission says the vast majority fall into the minimal-risk group, and it gives spam filters as the example. Whether a recommendation engine belongs there is a classification question, and the Commission's page does not settle it for personalisation. Treat minimal-risk status as a conclusion to reach with counsel, not as a default that saves the trouble of asking.
US attitudes lean cautious. Pew Research Center says about half of US adults are more concerned than excited about AI, and that concern among young adults is rising. Its survey summaries also find that views on AI and its pace of advance tilt negative, even among younger adults. Americans are open to AI on administrative and medical tasks, Pew reports, but they largely reject AI operating without human oversight. That last point is the most usable finding in the set.
The Commission makes a related point about explanation. It says AI decisions often cannot be explained, which makes unfair disadvantage hard to assess, and its example is hiring decisions. A personalisation system that cannot say why one customer saw one offer and another did not inherits the same problem, even though the Commission's example is different. The test is practical: can the team reconstruct the reason for a single output after the fact?
The US National Institute of Standards and Technology offers a voluntary route. Its AI Risk Management Framework, released on 26 January 2023, aims to build trustworthiness into the design, use and evaluation of AI. Its Generative AI Profile, NIST-AI-600-1, released on 26 July 2024, identifies risks unique to generative AI. NIST also says the AI RMF 1.0 is being revised as part of the White House AI Action Plan, so the current text may change. The difference between the two frameworks is enforceability: the AI Act is law in the EU, and NIST's framework is voluntary.
One Pew finding is a warning for anyone tempted to model customer sentiment with synthetic panels. Pew says synthetic 'digital twin' surveys estimate public mood poorly and are error-prone for some subgroups, including Republicans and Black adults. Using a synthetic panel to predict how customers will react to a feature is a substitute for asking customers, and the evidence here argues against the substitute.
Taken together, the regulatory and survey sources converge on two things: disclosure and human oversight. They do not show that personalisation itself lowers trust. That distinction matters for a board. A company can keep a recommendation engine and still meet the evidence, provided it tells people what is happening and keeps a person accountable for what the system does.
What the Netflix case shows
The most detailed personalisation case in the set is Netflix, and it is documented in a Harvard Business School blog post, not in research. The post says the recommendation engine uses viewing, search, scroll, ratings, time, date and device data. It says Netflix homepages were personalised from metadata tags, with 76,897 'altgenres' as of 2014. It also describes Netflix targeting House of Cards marketing at subscribers inferred to be fans of Kevin Spacey.
The House of Cards example is the one to study. The blog presents it as targeting based on inference, not on anything a subscriber chose to share. Whether customers found that helpful or intrusive is not in the sources. The case shows the mechanism, not the reaction, and a leader should not read a reaction into it that the evidence does not contain.
The same post gives scale. According to the blog, Netflix had 93.8 million subscribers worldwide as of December 2016, and about 100 million hours were viewed on Netflix daily as of August 2016. The blog does not link that scale to customer trust, so it sets the stakes rather than the outcome. It also notes that streaming services collect and own viewing data directly, while networks rely on Nielsen ratings as a proxy. Ownership of the data is part of why that precision is possible.
The blog also covers the 2006 Netflix Prize. A team won $1M in 2008, and the algorithm was never incorporated into the service. For a leader, that is a reminder that a model which performs well on a test is not yet a feature customers see and accept. Accuracy is one asset. Acceptance is another, and the sources do not show how to measure it.
Figures from the sources
Each figure as stated by its publisher. Pew figures are survey results on stated attitudes, not measures of customer behaviour.
- US adults who say AI-use disclosure in healthcare is extremely or very important (Pew Rese
- 72%
- US adults who say AI will worsen the patient-doctor relationship (Pew Research Center)
- 47%
- US adults who say AI will improve the patient-doctor relationship (Pew Research Center)
- 10%
- US adults who use chatbots for election information (Pew Research Center)
- 8%
Source: pewresearch.org; European Commission; digital.hbs.edu
What it means for a leader
Start with disclosure, because it is the one area where the evidence points somewhere. Pew's health finding, the Commission's chatbot rule and its labelling duties all say the same thing: people want to know when a machine is involved. In our experience the cheapest and most credible move is a plain line at the point where a recommendation or decision is made, not a clause buried in terms of service.
Use an explainability test before any feature ships. For each personalisation feature, ask whether you could state the inference behind it in one plain sentence to the customer it affects. If the answer is no, the feature is not ready. The House of Cards case is a reminder that inferred attributes can surprise customers more than the data they gave you, and surprise is the reaction a trust programme is meant to prevent.
Keep a named human owner for consequential decisions. Pew reports that Americans largely reject AI operating without human oversight, and the Commission's concern about decisions that cannot be explained points the same way. A person who can review and override the system is the simplest answer to both. Give that person a written list of what they may change, and review it each year, as you would any other delegated authority.
Check the rules before relying on any classification. Whether a system is high-risk, transparency-covered or minimal-risk under the AI Act is a legal question for counsel, not a product decision. The Commission says pre-market obligations for high-risk AI systems apply from 2 December 2027. Counsel should also confirm whether a firm selling into the EU falls within the Act. The sources point that way but do not settle it, and a US company should not assume it sits outside the rules.
Treat NIST's framework as a working structure, not a compliance badge. It is voluntary, and its text is being revised, so a company that adopts it should record which version it used and why. A short internal document that maps each personalisation feature to the framework's risk questions will serve a board better than a claim of conformity.
Limits of the data
The sources do not measure what this brief set out to measure. We found no peer-reviewed consumer-trust study of AI personalisation. The NBER items in the set are off-topic. The Harvard item is a course blog post rather than research, and the Pew items are survey summaries rather than peer-reviewed work. Survey answers record stated attitudes, not what customers do when a recommendation appears, and the Pew figures describe US adults only. Before a board acts on any claim linking personalisation and trust, a researcher should locate and read the peer-reviewed literature and check whether it says what this brief could not.