The debate over artificial intelligence has split the technology world into two camps that rarely seem to be describing the same technology.
On one side are those who warn that AI could pose an existential threat.
On the other are those who predict an era of unprecedented abundance and wealth.
For investors, businesses, and the public trying to make sense of it, the gap between the two can be overwhelming.
To better understand how to navigate this divide, we spoke with two experts who work with AI every day: Veni Dhir, Director of Corporate Venture Capital at ADP, and Ruchi Bhatia, Technical Product Marketing Manager at AWS.
Observe what AI can do now
Both the experts stressed paying attention to what AI can do now rather than worrying about what it can do in the future.
Veni Dhir said that while she has to price both the possibilities of existential threat and abundance, she was “not in the extinction camp as a base case.”
Dhir was speaking in her personal capacity as an investor and AI practitioner, and her views do not reflect those of her employer.
“The evidence that informs my view is deployment reality: AI systems are producing measurable productivity gains in specific domains — software development, customer support, document work — while hitting the same ceiling in the wild every time: edge cases, the need for human review, and the hardest last ten percent,” she added.
In her view, neither of the louder narratives is playing out on the schedule its supporters suggest.
“The abundance case is real but slower and more uneven than predicted; the extinction case rests on capability jumps that keep arriving later than the loudest forecasts said they would,” she said.
“I underwrite what I can observe in production, not timelines.”
Ruchi Bhatia also pointed out the distinction between present capability and future expectations.
“AI can already help people write code, summarize and research faster, and on some narrow tasks it matches or beats human performance.”
The expert pointed out that it doesn’t mean society is moving towards an “AI-powered utopia or mass unemployment in general. A lot of that debate jumps from what’s possible from a real capability point of view today to a large conclusion of what can happen 5 to 10 years from now.”
Why there is disagreement over AI’s long-term impact
The disagreement over AI’s long-term impact is not solely about what the technology can achieve.
It also reflects differing views on whether safety research, regulation, and governance can keep pace with its development, according to Veni Dhir.
She said experts can examine the same evaluations of AI systems and reach different conclusions because there are no reliable methods for predicting when broader capabilities will emerge.
However, Dhir argued that the divide also stems from differences in how people view institutions and their ability to manage technological risks.
“People who trust institutions see a manageable technology; people who don’t see a race to the bottom,” she said, describing the debate as a question of values that is often presented as a technical one.
Why AI capabilities do not automatically translate into economic gains
Ruchi Bhatia said the gap between an AI system demonstrating a capability and that capability being widely adopted is often overlooked when people assess the technology’s economic impact.
Adoption remains a significant hurdle, she said, noting that most organisations have not yet scaled AI beyond pilots.
Even when a system can perform a task effectively, businesses must consider whether they can afford to deploy it and whether it fits their operations.
“Running AI at scale still costs real money,” Bhatia said, pointing to the costs associated with running queries and generating text, images and videos.
“Cost also has to be weighed against quality, so there are trade-offs. Some industries have regulations that limit how freely AI can be used. Then there’s organizational change: are people ready for it? And is AI even needed for tasks that people already do well today?” she added.
Separate demonstrated results from forecasts and speculation
Bhatia said people assessing claims about AI should distinguish between 3 categories that include evidence of what has already happened, forecasts based on specific assumptions and speculation about possible future scenarios.
Evidence includes results from experiments, incidents within companies, labour-market data and productivity measurements.
For example, “a specific AI model outperforming another on a defined task,” the expert added.
She noted that forecasts, by contrast, depend on assumptions about how the technology might develop.
“If AI costs fall by a certain amount, adoption reaches a certain level and a share of tasks can be automated, what would the economic effect be?” she added.
Speculation goes further by exploring possible outcomes without presenting them as established predictions.
Bhatia said all three categories can be useful, but problems arise when they are presented with the same degree of certainty.
“The mistake is when these three get presented with the same level of certainty, and that’s when people get confused,” she said.
Headlines and benchmarks can distort perceptions of AI
Bhatia said headlines and benchmark results can contribute to misunderstandings about AI, particularly when strong performance on a specific test is interpreted as evidence that a system is ready for widespread use.
“A general model that performs well on healthcare overall may still struggle with a narrower, more specialized task that needs more domain-specific data. Strong benchmark results, even ones that beat human performance, don’t mean a model is ready for production,” Ruchi Bhatia added.
Bhatia said the gap between benchmark performance and practical application needs to be communicated more clearly, particularly as concerns about AI’s potential effects on jobs and society grow.
While public concern about technological change is understandable, she argued that people need more context about what AI systems can actually do and where their limitations remain.
The expert noted that the responsibility also lies with the people developing and deploying the technology to explain its capabilities and potential consequences, rather than allowing headline-driven narratives to dominate the discussion.
AI has the potential to improve a range of tasks, Bhatia said, but communicating how the technology could change existing processes is equally important.
How much of the AI debate is marketing?
Bhatia acknowledged that marketing plays a role in how AI capabilities and risks are presented, while arguing that claims made by companies can also be challenged and scrutinised.
“Claims can be challenged, and regulators have acted on false AI claims, so credible companies need evidence behind what they say,” Ruchi Bhatia said.
Veni Dhir said the safety warnings of AI models could be used to hype up the model’s capability.
“Safety warnings can function as a competitive moat (regulatory capture dressed as caution), and hype is how you raise the next round,” she said.
For investors and businesses, she recommended looking beyond public statements and examining what people are willing to stake their capital and careers on.
“When the people warning about risks are simultaneously raising billions to build the thing, take the warning seriously but discount the framing,” Dhir said.
She added that credible signals include predictions that can be tested against specific dates and statements that run counter to a person’s own commercial interests.
The post AI doom or AI abundance? Two tech insiders on how to tell evidence from hype appeared first on Invezz