The AI perception gap | WARC | The Feed
The Feed
Read daily effectiveness insights and the latest marketing news, curated by WARC’s editors.
You didn’t return any results. Please clear your filters.
The AI perception gap
Artificial intelligence has been a part of the mainstream conversation since the arrival of OpenAI’s ChatGPT in late 2022, but while the technology is becoming more powerful, users are also coming to terms with its limitations.
The crux of the matter
Generative AI is understood to have great potential if you need to produce a lot of text or imagery. Using a publicly available beta version, like the free ChatGPT that many people have tried, obscures the big costs of running an AI system. This is especially true when models are much less accurate than they need to be, as they don’t have robust oversight over what AI systems are producing.
Cycles of hype follow a pretty predictable curve when excitement dissipates once the rubber hits the road and a technology’s limitations become clear.
Research points toward impact-optimised image generation
First, the good news. Lumen reports positive results from a trial of AI creative tools with Diageo which could be optimised for “difference” performance metrics. In this case, the attention the ads commanded were “significantly” more than in Lumen’s average benchmarks.
For marketers needing to generate not ideas but sheer output rather than creative strategies or platforms, the signals are positive.
Or at least, positive for a balance sheet. Lumen also notes a 2023 Kantar study showing that on certain brand metrics AI had the potential to be at least as good as human-made creative. On the surface, this looks a lot cheaper at the HR level.
But costs are ramping up. You may be able to sack your creatives, but with demand soaring for AI-skilled engineers there may be a different set of costs to consider. Not least environmental, as the sheer amount of computing power increases.
Stanford AI index: money and nerves
Concerns about cost will be clear to readers of Stanford’s AI Index, which puts some numbers behind the staggering computing costs needed to train an AI – costs that are only rising as models get bigger. “OpenAI’s GPT-4 used an estimated $78 million worth of compute to train, while Google’s Gemini Ultra cost $191 million for compute,” notes the AI Index report.
There is, admittedly, a lot of money flowing in. Funding for GenAI has octupled since 2022 to reach $25.2bn. But investors are not the public, 52% of whom express nervousness about AI coming to products and services – a 13% year-on-year increase.
With figures so large required, industry leads the academic or government in development by a long, long way. But industry brings with it challenges: companies need to make sure their models look (and, vitally, sound) impressive to clients and imposters. Robust, standardized evaluations of LLM responsibility are scant.
AI is going to take longer to generate returns than thought
Data from ETR, a research firm specialising in an ongoing panel of tech decision makers, finds that gaps in the rate of application of this new technology are emerging, despite the majority (80%) of companies surveyed having evaluated the potential applications of AI in their business.
Unsurprisingly, leadership tends to be much more optimistic than the experts, with some of the biggest gulfs found in the expected short-term returns (three months) – a timeframe in which 21% of C-suite leaders expect returns compared to just 10% of experts. This is based on a smaller subset of respondents at companies that have “implemented” generative AI features.
“I do think the hype cycle, because of how quickly it moved, the pendulum is stalling and swinging back a little bit,” explained Erik Bradley, ETR chief strategist and research director, in comments to Sherwood, a business news website. “People are recognizing it's still important, it's still going to be there. It might not be the panacea, the savior of their enterprise at this stage.”
Sourced from Lumen, WSJ, WARC, Stanford, Sherwood
Email this content