Few readers today can ignore artificial intelligence and the impact it has had on publishing, and whilst the technology may prove to be a remarkable tool, readers have struggled to come to terms with its presence. In games criticism, an increasing number of websites have resorted to AI-generated write-ups, texts produced by language models, and the irritation these have caused amongst gamers has, in many cases, turned into something bordering on intolerance.
People ultimately want to know what other people think. This conclusion is by no means far-fetched at a time when human interaction is taking on renewed value, precisely because our attention is absorbed by phone screens, and conversations, to use the well-established English phrase, are dwindling. An AI-generated review may provide a contribution, a summary of features or a list of system requirements, but it rarely stands up to comparison with a detailed article written by a talented critic with extensive experience in the field.
Whether the reader is a keen player of online blackjack or prefers PC games, forming an opinion about a game based on an AI-generated article is, for this reason, not recommended. Public unease regarding machine-generated content has, in the meantime, reached both the comments sections and the agendas of regulatory authorities. Readers’ patience is wearing thin, particularly amongst those gamers who continue to refuse to read content produced exclusively by a chatbot.
Only genuine opinions
Gamers form a discerning community, whose members rely on one another for feedback and practical advice; consequently, what is valued is the actual gaming experience, and a review pieced together from material taken from other authors contains very little of this. Such a text tends to lack originality and, in the worst cases, is a haphazard jumble of statements that do not form a coherent argument. When a summary is compiled from numerous online sources, the reader has no way of determining whether the text is accurate, up to date or even dedicated to the game in question. Artificial intelligence is, to use the well-established English expression, prone to hallucinations, that is, it is prone to hallucinations, and readily repeats information that has become outdated elsewhere; above all, it cannot play the game.
Games, a subjective hobby
A game that one player finds downright off-putting may captivate another for hours on end, which means that gaming is a subjective experience and that any review speaks from a particular gaming experience. No AI tool can explore a game in all its depth; what it produces is a recombination of other reviewers’ work, presented anonymously, which, accordingly, carries little weight. Respected critics have earned their living through knowledge accumulated over years of engagement with the industry, and readers value them precisely for the individual perspective that a language model simply cannot replicate.
Generic descriptions are of limited value
Another shortcoming of AI-generated writing is the lack of detail. Many AI-generated articles consist largely of generic descriptions taken from other sources, so the reading experience is generally monotonous. Readers looking for thought-provoking material that cannot be found elsewhere are poorly served by such texts.
Understanding the community matters
The gaming community has developed its own slang and terminology, accompanied by a body of shared references that only a member of the community can fully understand. A passionate human author is therefore better placed to write a text that sounds natural to a particular gaming community than a formulaic review produced from a template.
Authors have honed their craft over many years
The following concern relates to all areas of AI-generated content. Writers who have honed their craft over years of work, largely spent in newsrooms, do not deserve to see their jobs taken over by software that lacks both personality and human emotion. For this reason, many readers refuse to contribute to the erosion of the profession by paying attention to AI-generated content.
What the research on credibility reveals
The intuition that people want to know what others think has a long history in social psychology, and at Yale, in the early 1950s, it was put to the test in an experiment. In 1951, Carl Hovland and Walter Weiss published a study in *Public Opinion Quarterly* in which history students were given identical articles on four topics, including nuclear submarines and the future of cinemas. Each article was attributed either to a highly credible source, such as the physicist Robert Oppenheimer or *Fortune* magazine, or to one in which the students had no confidence. Opinion shifted towards the author’s position by an average of 23.0 per cent when the source was credible and by only 6.6 per cent when it was not. A difference of 16.4 points after a single reading is considerable. Two years later, in *Communication and Persuasion*, Hovland, Irving Janis and Harold Kelley broke credibility down into two components, which have since shaped research: expertise – that is, the public’s belief that the source knows the subject – and trust – that is, the belief that the source will truthfully report what they know.
A less frequently cited finding from the same study relates directly to the debate surrounding AI-generated reviews.
Students retained the factual content equally well from both types of source, answering 84.0 per cent of the questions correctly after reading a credible source and 81.5 per cent after reading an untrustworthy one – a difference the authors considered insignificant. Credibility therefore determined whether the conclusion was accepted, whilst the information was assimilated in both situations. Applied to games criticism, this finding suggests that the argument outlined above needs to be qualified in one respect. An AI-generated text can report the frame rate, price, campaign duration and list of platforms just as well as a human critic, and readers will glean this information from it. What it cannot provide is a verdict that the reader has reason to accept, since the expertise on which a verdict is based, in this field, is the experience of having played the game.
The model has limitations that need to be clarified. The subjects were American students reading printed articles in a classroom setting seventy-five years ago. The credibility effect has also proved unstable over time. After four weeks, agreement with credible sources had declined, whilst agreement with non-credible sources had increased, until the two had become almost identical; Hovland’s research group had already termed this pattern the ‘sleeper effect’. Subsequent experimenters, including Patricia Gillig and Anthony Greenwald, have struggled to reliably replicate this delayed reversal, suggesting that it occurs only under certain conditions. Credibility is, moreover, an attribution made by the reader and can be faked. A review written by a machine and published under the name of a fictitious human author lends itself a credibility it has not earned, and the problem is more serious than in the case of an AI-generated abstract that is openly labelled as such. Credibility borrowed in this way is difficult for any reader to detect from the text alone.
Evidence of authorship, the reader’s test
Building on this distinction, the more useful question for the reader concerns the evidence of authorship contained within a review, as the flaw identified above predates artificial intelligence. Long before language models, content farms were publishing reviews rewritten from press releases and material from other publications, by authors who had never even installed the game in question. A human byline did not protect readers from those texts. Conversely, a reviewer who has spent sixty hours with a title and then uses software to polish their phrasing has lost nothing of the experience on which the review is based.
Regulatory authorities and search engines have arrived at a similar criterion. The Federal Trade Commission’s rule on consumer reviews and testimonials, announced on 14 August 2024 and in force from 21 October 2024, prohibits fake reviews, including, explicitly, those generated by AI and those written by people with no experience of the product. The rule allows for civil fines of up to nearly $52,000 per violation, an amount adjusted annually for inflation. It targets reviews and testimonials used in commerce and does not regulate editorial criticism, but the test it applies coincides with that used by gamers. Google’s guide to writing reviews, last updated in December 2025, requires reviewers to provide evidence – such as images, audio recordings or links – to document their own experience with the product, and to present quantitative measurements of its performance.
In practice, such evidence takes easily recognisable forms. The first is the number of hours played, along with the platform used. The second is the patch or version reviewed, as games change considerably after launch: updates rebalance weapons, fix save systems and, at times, overhaul entire game modes, meaning that a verdict given in the first week may be outdated a month later. Other elements include original screenshots or gameplay footage, rather than material provided by the publisher; a clarification as to whether the copy was provided free of charge by the publisher; and specific moments, bugs or sudden spikes in difficulty that only someone actually playing the game would have encountered. A review that contains none of these elements remains unverifiable, regardless of who wrote it or what it says. Most of the items on this list cost a reviewer who has actually played the game nothing at all.
How to find trustworthy reviewers
When it comes to discovery, the reader faces a structural difficulty. Search results and social media feeds present the text before the author, so that a review is usually encountered without the body of work that would allow its credibility to be assessed. Expertise and trust, in Hovland’s terms, are qualities that a source demonstrates over time, through a consistent track record, a history of judgements that have subsequently proved sound, and reliable information about who owns the publication and how it is funded. The owner is usually listed on the ‘About Us’ page or in a legal notice, and the absence of this information speaks volumes in itself. Consistency is the easiest indicator to verify, as the reader can look at a reviewer’s last ten reviews to see whether the ratings and reasons given are consistent with one another.
Funding deserves special attention in the press covering casinos and real-money gambling, where affiliate commissions are common, and a review may, in fact, be paid for by the very operator it is evaluating. A clear disclosure of such arrangements forms part of the evidence that the reader is entitled to see.
The same distinction between curated and reviewed content also applies to website organisation. Aggregator directories automatically compile listings from data feeds and information sourced from other websites, whilst editorial directories only accept a site after a person has reviewed it and written a description of what it offers. The difference between aggregator directories and editorial directories reflects, with a certain degree of accuracy, the difference between a compiled review and one written by a critic. A curated directory can therefore help the reader to find publications whose existence, ownership and subject area have been verified by someone, and to return to them later; and this is precisely how a reader comes to know a source over time.
Language models face a related problem from the opposite direction, as they too must decide which sources to treat as reliable, whilst structured, human-verified records have begun to function as curated reference points for AI systems. The proposed remedy for the machine’s lack of reliability is, in fact, the same human verification that readers demand of reviewers. The reader and the model ask a source the same question, namely whether anyone has verified it.
The limitations of such a directory must be clearly stated. An entry in a directory confirms that a publication exists, who runs it, what it covers and where it can be found. It does not guarantee the soundness of any particular review. It is not a recommendation to buy a game, nor does it reveal anything about a critic’s taste. For real-money casino games, the decisive verification takes place elsewhere. Real-money online blackjack is restricted to adults, generally those aged 21 or over, in US states where it is legal. An operator’s licence can be verified in the register of the relevant state’s regulatory authority – which in New Jersey is the Division of Gaming Enforcement – and no editor can substitute for this verification.
Judgement as a product of experience
Thus, gamers’ resistance to AI-generated reviews reflects a demand older than any machine: namely, that a judgement should be attributable to someone who was present when the matter in question was assessed. Experiments at Yale have shown that facts circulate independently of their source, whereas beliefs do not; and game criticism, which operates almost exclusively through judgements, is the field in which this asymmetry is felt most acutely. It can be concluded that a review worth reading – whoever or whatever may have contributed to its writing – is one that bears the traces of the game within it, and the reader’s task is to learn to recognise them.
