Andy Jassy, chief government officer of Amazon, speaks throughout an unveiling occasion in New York, US, on Wednesday, Feb. 26, 2025.
Bloomberg | Bloomberg | Getty Pictures
Anybody who has ever gone in quest of a product evaluate on Amazon is aware of how priceless the expertise of different customers could be, and the way simple it’s to fall down the rabbit gap of buyer feedback, from five-star raves to one-star takedowns — typically tons of of phrases to get to the purpose. As Amazon continues to roll out AI options it says will make procuring simpler, AI-generated audio descriptions of merchandise are within the combine and scraping that buyer commentary — perhaps finally settling right into a place to interchange it as a go-to supply of shopping for data.
Referred to as “Hear the Highlights,” the AI-voiced product descriptions depend on a big language mannequin to script the abstract based mostly on a wide range of sources, pulling from Amazon’s product catalog, buyer opinions, and data from throughout the net, after which translating the content material into short-form audio clips. The summaries started rolling out throughout the summer season on choose merchandise to a subset of U.S. prospects and have now reached all U.S. prospects as a button within the cellular procuring app, scaling as much as cowl over a million merchandise.
A novel enchantment of Amazon, and different e-commerce choices, has at all times been the flexibility to get data from precise customers, not simply product descriptions. In fact, bogus opinions have lengthy been an issue for Amazon, despite the fact that it bans use of paid promotion and different inauthentic types of evaluate writing. You continue to need to type via varied efforts to sport the system, and in recent times, indicators that the most recent evaluate writers are chatbots like ChatGPT. However precise folks sharing their precise, idiosyncratic expertise with a product — prospects as a supply of data, and knowledgeable resolution making — have been a key a part of the educational course of from clothes and shoe sizes to security and extra slim questions: When buying a brand new toaster, does it do bagels, is the timer setting correct, how simple is it to scrub the crumb tray?
Can AI enhance on that? Opinions are by nature unwieldy within the mass individuality, however human readers have confirmed fairly adept at distilling what they want from the human chaos. In some sense, the mind of the common Amazon evaluate reader is a fairly good giant language mannequin, skimming and choosing out key phrases and key data, so an AI should do it at the least in addition to a human would if it is so as to add to the shopper expertise as promised.
AI and cognitive overload
AI does have its benefits. For one, it will not expertise cognitive overload when dealing with an electrical tea kettle on Amazon with over 32,000 opinions. It will probably comb via the info, however can it provides us solely what we’d like or wish to know? That may nonetheless be tough for AI to get proper, and it might be inherently problematic to combine product catalogs, buyer opinions, and net data right into a single distillation – with the sources of this data every coming with its personal set of intentions.
“It is essential to acknowledge the place AI is presently sturdy, reminiscent of in automation and sample recognition, and the place it nonetheless falls brief, like in judgment-heavy duties,” stated Ankur Edkie, CEO & co-founder, Murf AI, which creates AI voiceovers. “A key query is whether or not there is a method to consider buyer context as an enter whereas producing these summaries,” he stated.
The worth of AI, in response to Edkie, is discovering the correct problem-capability match. If that is not achieved, the sense of gimmickry is more likely to sneak in via a door left open for AI fatigue, which he says shoppers are doubtless experiencing by now.
Amazon’s Hear the Highlights AI-audio summaries are presently the identical for each person.
“The AI abstract must seize nuance and context. For instance, even just a few unfavourable opinions on security can outweigh many positives on different options,” Edkie stated, including that if a buyer is targeted solely on product efficiency, then summaries that emphasize worth may not be related.
The flexibility to consider context, and to make the evaluate course of extra of a dialogue between shopper and machine is probably going the place the know-how is headed – in different phrases, towards the agentic facet of AI, the place Amazon can also be actively including to its AI commerce instruments, reminiscent of Rufus, and a procuring instrument referred to as Pursuits AI, which prompts customers to explain an curiosity “utilizing your personal phrases,” after which it generates a curated choice of merchandise. That function, rolled out within the spring, is separate from the primary search bar on Amazon’s web site.
Chat and audio summaries will stay among the many methods to have interaction, however having real-time conversations with an AI voice agent — asking particular questions, clarifying considerations, and getting deeper insights from opinions — is what is going to shift the expertise from one-way supply to two-way discovery, making it much more personalised, in response to Edkie. “At present, you may work together with Rufus via textual content or by utilizing voice enter, however Rufus can’t speak again, its responses are text-only,” he stated. “With voice brokers, nonetheless, you may have a two-way dialog with a bot that speaks to you,” he added.
For now, one section of shoppers more likely to see instant advantages is visually impaired customers, making accessibility an intriguing side of the function, however the voice have to be high-quality and ship the content material precisely.
Brian Numainville, principal at shopper analysis agency Suggestions Group, says by offering an audio-based different to visually introduced data, all these options have the potential to make procuring extra accessible, changing detailed textual content into simply consumed audio summaries. Nonetheless, for it to really profit folks with visible impairments, the function have to be thoughtfully designed. In accordance with Numainville, this would come with making certain full compatibility with display screen readers and keyboard navigation, offering clear, structured, and concise summaries, and avoiding overly lengthy or complicated audio displays. The standard and readability of the AI-generated voice may even considerably influence usability.
“The shift from numerous human opinions to AI-generated summaries would possibly imply dropping essential nuances, context, and private touches,” Numainville stated.
Threat of dropping distinctive shopper insights
The tendency of AI to concentrate on widespread themes can dilute responses even because it distills them.
Human opinions have a tendency to incorporate digressive tales and particulars round extremely particular use circumstances — consider what motivates somebody to write down a evaluate within the first place, and the way this pairs effectively with the consumer’s anxieties and decision-making course of — all of which will help to make the procuring expertise extra private and insightful.
“AI would possibly overlook distinctive insights or area of interest wants that do not align with nearly all of responses,” Numainville stated. “Moreover, the flexibility to critically interpret opinions — like recognizing biases or trusting sure reviewers — is diminished with AI summaries.”
Scraping a substantial amount of content material to generate the summaries additionally sacrifices the annotation that may distinguish amongst product descriptions, buyer opinions, and net data.
“It is not 100% clear how a lot of the product description versus opinions is used within the present implementation — Amazon hasn’t absolutely detailed this combine,” stated Numainville.
Amazon declined to remark, referring CNBC to publicly accessible data on the function.
“Product descriptions are sometimes marketing-driven and emphasize positives, whereas opinions mirror actual person experiences and embody each execs and cons,” Numainville stated. Combined with out attribution — which by definition flattens data sources — it might be troublesome for purchasers to discern ad-speak from real opinions, with the outcomes be one thing like native promoting.
“This mixing, if it occurred, may unintentionally mislead shoppers by lending factual authority to subjective opinions or disguising promotional materials as unbiased opinions,” Numainville stated.
Analysis does present that prospects are inclined to belief voice as a format of data supply, no matter how balanced the data being delivered is.
“It appears more likely to me Amazon’s want to promote merchandise would weigh extra extremely than incorporating all vital views on a product,” stated Tama Leaver, professor of web research at Curtin College in Australia.
One other concern is that “AI would possibly weigh common and general scores, whereas consumers usually take a look at the few unfavourable opinions – the one stars – even when there is not a variety of them,” Leaver stated.
Dr. Nauman Dawalatabad, a analysis scientist at Zoom Communications, stated in his private view the know-how is shifting within the course of higher buyer expertise. “I take it as know-how serving to us to make knowledgeable choices,” he stated, citing the cognitive fatigue and wasted time that may come from studying via buyer opinions.
If a streamlined description finally ends up resulting in impulse purchases and regrets, that’s on the customer, he says, and no completely different than the way in which issues have at all times operated — the identical fees of refined coercion might be made towards all advertising efforts. He thinks that as voice-based agentic AI continues to evolve and shoppers begin speaking (as a substitute of typing and looking out) with an AI agent and describe what they need, “it’ll get you precisely what you want.”