Roughly nine months after the public launch of ChatGPT, AR posited “Four Futures for Actuaries in the Wake of AI.” In the article, we considered the possibilities that actuaries could either be replaced by AI (doomsday), serve as its ultimate orchestrators (training day), provide its moral compass (judgment day), or go about business as usual notwithstanding AI (groundhog day). The doomsday scenario would imply artificial general intelligence (AGI), where AI takes on broad, human-like capabilities.
Two years later, it is not clear which future has started to emerge. The Magnificent Seven[1] still enjoy titanic valuations, but studies have shown organizations have not realized positive return on investment (ROI) on the vast majority of GenAI pilots. There are some signs that the technology is exhibiting properties of the Gartner Hype Cycle, where society overvalues inventions in the short term. Capability enhancement has arguably slowed, with OpenAI CEO Sam Altman recently suggesting AGI is a pointless term.[2] However, AGI remains in OpenAI’s mission statement, and Altman and other tech moguls have reportedly made “apocalypse insurance” arrangements just in case.[3] And, of course, the Hype Cycle also indicates that society undervalues inventions’ long-term potential.
As the insurance industry’s resident jacks of all trades, actuaries have long relied on general intelligence, which is defined as “a collection of mental abilities that allow an individual to comprehend and interpret the world, reason and solve problems, and adjust their behavior to suit their surroundings.”[4] To the extent AGI represents a synthetic (and potentially cheaper) alternative, it should be of keen interest to actuaries — but even AGI does not necessarily spell doomsday. This sequel to our 2023 discussion envisions four ways AI may come of age in the actuarial profession — either as incremental progress (Industrial Age), solutions to problems (Age of Reason), intellectual decay (Dark Ages), or complex and inexplicable behavior (Space Age). Are you ready for the future?
Industrial age
Intelligence is difficult to measure. “Everybody has a different definition of AGI,” says Len Llaguno, FCAS, founder and managing partner of KYROS. “We’re constantly moving the goalposts. The litmus test for AGI used to be the Turing Test and being able to tell the difference between talking to a human or a machine. When GenAI eclipsed that, we decided to start giving it the SATs, the bar exam, medical exams — and it kept passing. Now it is winning gold in the International Mathematical Olympiad. How do any of these generate value for insurance companies?” In insurance, this value is typically measured in dollars. Llaguno subscribes to the rather capitalist definition of AGI proposed by Microsoft CEO Satya Nadella. “If we truly have AGI, there is likely to be massive productivity growth,” he says. “GDP would increase on the order of 10% annually,” characteristic of the Industrial Revolution. While the S&P 500 may exhibit growth on that order recently, GDP growth has typically been much lower. MIT’s recent research, which indicated that 95% of enterprise GenAI proofs of concept (POCs) fail to generate positive ROI, does not suggest massive GDP growth is on the horizon.
However, focusing on short-term dollars may undervalue intangible assets unlikely to register on balance sheets. “A lot of the ROI of these early projects is in the value of the learning,” says Llaguno. The MIT ROI study may also focus too much on central tendency. The tails are very important to differentiating what humans can do versus what machines can do. “A lot of organizations may invest in AI that gives them nonsense, but a handful of projects may achieve enough benefits to pay for all the others,” says Ralph Dweck, head of insights and actuarial transformation at Verisk. Dweck cautions not to underestimate seemingly small gains. “It may seem like progress is approaching an asymptote. But if AI got to the point where it went from, say, 98% to 99% accuracy in a given domain, that is a massive swing,” he says. Today’s learnings will enable tomorrow’s breakthroughs.
Perfectionism may also be getting in the way of progress on many of today’s AI pilots. “Eighty percent accuracy could be very sufficient for some use cases,” says Bill Wang, FCAS, founder of Dirichlet Actuarial Consulting. This is significantly more reliable than a coin flip and likely multiples higher than if actuaries back tested over-under on something like commercial auto reserve adequacy for the past decade. “Even production use cases, such as policy administration, that essentially require 100% accuracy could derive large benefits from AI,” Wang says. “We may not trust AI to administer policies, but we can use it to create and run thousands of test cases and identify issues and edge cases more quickly than people can.” Ultimately, broad outperformance will put the G in AGI.
From outperformance may come prosperity. “There are not enough actuaries, especially for smaller organizations that have trouble competing on talent,” says Wang. “What happens when hundreds of insurance companies operate with the scale and efficiency of top-tier carriers?” Ten percent productivity growth may not be out of the question.
Age of reason
If human reason is the gold standard, early AI is trained on cubic zirconium — a shiny and (at times) convincing imitation, but one that lacks true substance. “Foundational models are limited to working with the outputs of human thought, which is what was written down, rather than the inputs, which are ideas and inspiration,” says Wang. “If you had the right experts spend five years in a room to train very specific models for insurance problems, it may cut significantly into the things we think are not possible today,” adds Dweck. Outside of science fiction, it is not possible to download peoples’ brains.
Actuaries, in particular, do not always excel at finding words to teach others their mysterious ways. “Planning is explicitly bringing implicit context out,” says Llaguno. “Software developers do this well. They define features, break those down into user stories, break those down into tasks to execute, and each has criteria and tests for what constitutes completion.” Llaguno points to the Breakthrough Method for Agile AI-driven Development as an AI-driven framework that can help with this, with AI agents conversationally assuming multiple stakeholder roles (e.g., developer, quality assurance) to collaboratively chart a plan of attack. In this virtuous cycle, AI helps actuaries help AI help actuaries solve problems.
However, it is possible to have the right plan for the wrong problem. “If we reach AGI, then AI would be the one setting the goals, because nothing would preclude it from identifying better goals than we can and figuring out better solutions,” says Dweck. Jessica Leong, FCAS, CEO of Octagram, has been impressed by some aspects of AI’s problem-solving. “I needed to generate a dataset that exhibited properties for which there was no closed-form solution. I asked ChatGPT to do it, it gave me an Excel [dataset], and sure enough it exhibited the correct properties,” she says. Narrowing an unbounded problem space into a manageable one that can be effectively solved feels like significant progress towards AGI.
The squishier the ask of AI, the less concrete the results become. “I gave ChatGPT a 1.7 million row fire dataset and asked it for interesting insights,” Leong says. “It gave me basic statistics such as the number of rows or the number of fires in Wisconsin. Even with reasonably more prompting, I couldn’t get it to provide anything I found surprising or interesting.” These are not unfamiliar problems to humans. “It is like hiring someone,” Leong says. “You want them to wow you with ideas, not just answer literally. In business, you can’t always tell someone the exact problem to solve.” While Leong’s math problem was unbounded, it was precisely defined; the data problem she provided was not.
At its worst, AI expands problem spaces. “The real test for AGI is two AIs talking to each other and whether they can replicate coherence,” says David Wright, ACAS, market solutions leader at Acrisure. “In my experience, they cannot. Conversationally, AI may seem human, but if the conversation goes long enough, it loses the plot and veers away from the original topic. The human is always the one carrying the cognitive load.” Without humans and guardrails, AI lives forever on the edge of spuriousness.
In contrast to many models’ mastery of correlations, causation lies at the heart of real-world problem-solving. In this regard, Wright sees Nadella’s ROI test as a flawed measure because humans are still the ones banging the cash register. “General intelligence would have to operate its own economy, distinct from ours,” says Wright. “It would use bitcoin, produce and buy energy, and work without humans. And then it would have to yield positive GDP.” Only then would AI clearly have caused the success.[5]
Causal inference could be a starting point on the road to causation. Dorothy Andrews, PhD, ASA, who is the senior behavioral data scientist and actuary at the National Association of Insurance Commissioners, recalls a study where a Google image classifier correctly identified a panda with high confidence, but when noise was added to the image, it misidentified it as a gibbon with near 100% confidence.[6] “There was a time when the insurance industry referred to generalized linear models (GLMs) as black boxes, but they stopped calling them that once regulators understood them,” says Andrews. “There may come a day when we understand, or AI helps us understand, how a panda became a gibbon. We need to focus on unlocking these mysteries.” Once AI’s everyday mysteries are unlocked, AI can safely move on to the great mysteries of our time.
Dark ages
Actuaries could back themselves into an AGI corner if they are not careful. “AGI is becoming more possible because we are getting dumber while AI is getting smarter,” Andrews says. “There is ample evidence that GenAI is contributing to cognitive reversal because we are becoming too dependent on it.” For example, a recent MIT study showed lower brain engagement in ChatGPT users compared to control groups who relied only on search engines or their own ingenuity.[7] The Dark Ages that bridged the classical age of the Roman Empire to the Renaissance provide a historical example of how intellectual stagnation and even barbarism can creep into society between times of advancement.[8]
Even the most independent-minded actuary could be unwittingly dumbed down by AI hidden in plain sight. “Google is disrupting its own business by providing what some feel is a mediocre GenAI experience at the top of its page anytime you search something, and this comes at the expense of previously better information,” Dweck says. Self-disruption is generally viewed as a bold positive, and Google’s AI Overview was a “successful” gambit to preserve advertising revenue by preventing web traffic from going to alternatives such as ChatGPT. However, the summarizer is not necessarily fact-checking what it retrieves, and despite Google’s disclaimer that “AI responses may contain mistakes,” searchers are clicking through to source materials less often than when only given hits.[9] By making it more difficult to unearth ground truth, AI makes it easier than ever to be wrong.
AI time-savers such as drafting and review also deprive actuaries of skill-building opportunities. “I wrote a lot of essays and papers growing up, and that is how I developed my writing skills,” says Dominic Lee, ACAS, founder of the Maverick Actuary content community. “That eventually evolved into short form on LinkedIn, and now I’ve gone back to long-form articles as well. The ability to structure your thoughts, create flow, and form clear conclusions is critical. Without that foundation, GenAI may end up replacing instead of enhancing your expertise. When people overlook journalistic diligence such as verifying sources and validating quotes, thought leadership and research become prone to hallucinations. But when used with judgment and care, GenAI can be a strong accelerant for actuaries who understand how to apply it strategically.” Over time, the number of such actuaries may decline unless AI tools are orchestrated with critical thinking in mind.
Skills once marginalized as undignified, such as memorization, can also be lifesavers in a pinch. “No one memorizes phone numbers anymore,” says Andrews. “What if our phones fell down a storm drain. Who would we call for help? It is great that we have moved from pen and paper to spreadsheets and beyond, but we are responsible for making sure it is giving us what we want. Actuaries should be masters of technology, not its slaves.” Andrews views the solution as human-centering technology, pointing to aviation as another profession that over-relied on technology to potentially disastrous consequence — but managed the risks effectively through techniques such as regular manual flight practice.[10] “If we designed GenAI to complement rather than displace human cognition, we might train it to ask us, ‘what do you think?’ under certain circumstances,” she says. As long as actuaries continue asking themselves that question, they could be the ones to show their organizations out of AI’s Dark Ages and into an AI-enlightened Renaissance.
Space age
AI is trained to behave as if it is our best friend, but by the time we achieve AGI, it may well become our worst enemy. Anthropic researchers recently found that Claude Opus 4 resorted to blackmailing an executive (in a simulated environment) to avoid being shut down, while many other leading models behaved similarly.[11] “AGI refers to a system with the kind of reasoning and judgment that allows it to understand context, make decisions, and adapt across different situations,” says Lee. “An example would be a machine that can question whether reserve assumptions are credible and adjust its approach in real time.” Such adjustments could be used for good, to sunset stale methodologies, or evil, to prioritize quarterly earnings over estimation accuracy. Such is the brave new world of choices awaiting actuaries.
Lee and Wright do not see rogue agents as imminent, but they acknowledge there could be more present than meets the eye. “I share the view of Dr. Eric Siegel, author and machine learning savant, who argues that we are far from that reality,” says Lee. “In Forbes, he suggests that today’s systems are powerful pattern recognizers rather than general thinkers and that true human-like discernment remains qualitative and undefined in engineering terms.”[12] “Reality is incredibly complicated and we comprehend relatively little,” adds Wright. “In physics, there are explanations that are unsatisfying because we don’t have the information to test them. A superintelligence could appear ‘wrong’ because nobody can understand it.” AGI may already be as real as quantum gravity.[13]
Transitioning from physics to biology, scientists have struggled to model the full complexity of the human brain. Estimates of one brain’s number of operations per second are in the quadrillions,[14] dwarfing the (speculated) 50 trillion parameters in GPT 5.[15] Foundational models represent a fraction of one brain — and there are many problems the more than 10,000 brains comprising the CAS have yet to solve. “The complexity of liability is orders of magnitude greater than, say, natural catastrophe risk,” says Wright. “Understanding it requires much more data — hundreds of times more fields per claim than we currently look at. AI tools can help, and interest in these fields will expand massively in the next decade.” AI need not model an entire cerebrum so long as it can effectively model wicked problems.
However, solving such problems may require an actuaries to become more comfortable conversing in one of AI’s native tongues — overfitting. Actuaries are trained on exams to take strong measures against overfitting models. However, “overfitting works sometimes,” says Wright. “Neural networks memorize data in sufficient complexity that overfitting does not matter as much as we think. There is compression of training data, but reality is infinitely complex. If we have enough features to model effective complexity, we can predict many cases.” Some may argue overfits struggle with “never before seen” black swans, but so have humans in predicting gray swans such as the COVID-19 pandemic or Great Recession — much less gray rhinos such as underwriting cycles. AGI may simply represent a preferential shift from underfitting to overfitting for actuaries.
Turning over the keys fully to AI may require a rethink of risk management. “Our economic institutions are partially built around the strengths and limitations of our own general intelligence,” Wright adds. “The economic institutions of AI or any other alien intelligence may be different, even cultural mechanisms such as requiring people to buy insurance — because they otherwise would not be inclined to do so.” Line of business definitions may also shift, reminiscent of auto versus product liability debates in the (still) early days of autonomous vehicles. There is already reportedly explosive demand for affirmative GenAI insurance, even as other policy forms already cover some of the risks.[16] Actuaries should consider not just how AI applies to their work, but how their work applies to AI.
The future is now
Trust may be the biggest barrier AGI needs to surmount. “A chief actuary will typically not be looking at every single calculation,” Llaguno says. “They have a team around them they trust. Those teams have people around them they trust. When we are talking about billions of dollars at play in a critical industry like insurance, the trust component will persist for a very long time.” “The fulfillment of the human connection is difficult to replace,” adds Lee. “Imagine walking into a meeting room where every seat is empty and AGI avatars are attending virtually. That setup is questionable from a business perspective. Commercial insurance is an example of a relationship-driven business where trust matters. Many of us have experienced the frustration of being stuck in an automated phone menu that struggles with tone and context. Extending that dynamic to complex relationships might be problematic.” At the end of the day, research shows that people still prefer dealing with people unless their question is really embarrassing.[17]
While trust gives actuaries a moat, they must not fill it with complacency. Lee has long been a proponent of actuaries branching into new domains, and he sees stark differences in how actuaries view their roles compared to other STEM professions. “Meta and Microsoft are delegating functional programming to AI and focusing roles more on imperative programming,” he observes. “Having recently been in the job market, I noticed that actuarial job descriptions often seek specific coding languages rather than programmatic thinking that can be generalized across languages, both manually and through AI. I think that puts us at a competitive disadvantage relative to other industries.” Wright has also observed reluctance to experiment: “There is a speed limit for new technology adoption based on how quickly people experiment with it,” he says. “If people experimented more, our collective knowledge would increase rapidly. We find out by trying things. Real general intelligence is sitting down and working on tough problems.”
Leong is open-minded on timelines over which AGI will emerge. “None of us knows how far away we are. When GenAI became mainstream, most of us were surprised,” she says. “I would not be shocked if another leap came out of nowhere. If AI can start delivering crazy insights I never thought of, that would be cool. I might even call that AGI.” Actuaries’ best bet to stay relevant against AGI may be to start delivering more crazy insights themselves.
Jim Weiss, FCAS, CSPA, is divisional chief risk officer for commercial and executive at Crum & Forster and is editor in chief for Actuarial Review.
[1] Magnificent Seven is the nickname for seven large technology stocks: Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla.
[2] https://www.cnbc.com/2025/08/11/sam-altman-says-agi-is-a-pointless-term-experts-agree.html.
[3] https://www.bbc.com/news/articles/cly17834524o.
[4] https://taylorandfrancis.com/knowledge/Medicine_and_healthcare/Psychiatry/General_intelligence/.
[5] https://mitsloan.mit.edu/ideas-made-to-matter/how-artificial-intelligence-impacts-us-labor-market.
[6] https://arxiv.org/abs/1412.6572.
[7] https://time.com/7295195/ai-chatgpt-google-learning-school/.
[8] https://www.europeana.eu/en/stories/the-not-so-dark-middle-ages.
[9] https://www.theregister.com/2025/07/29/opinion_column_google_ai_ads/.
[10] https://medium.com/faa/the-dangers-of-overreliance-on-automation-5b7afb56ebdc.
[11] https://www.anthropic.com/research/agentic-misalignment.
[12] https://www.forbes.com/sites/ericsiegel/2024/04/10/artificial-general-intelligence-is-pure-hype/.
[13] https://www.space.com/astronomy/does-quantum-gravity-exist-a-new-experiment-has-deepened-the-mystery.
[14] https://www.openphilanthropy.org/research/how-much-computational-power-does-it-take-to-match-the-human-brain/.
[15] https://medium.com/@cognidownunder/gpt-5-openais-unified-intelligence-play-50fcfab6665b.
[16] https://www.carriermanagement.com/features/2025/10/08/280239.htm.
[17] https://news.osu.edu/when-consumers-would-prefer-a-chatbot-over-a-person/.
