Reconciling AI Risk and Opportunity: A Net Assessment for Investors

408 Pages, Two Columns: What Anthropic’s S-1 Reveals

Investors are poring over Anthropic’s S-1 prospectus in anticipation of the largest IPO in history. The document exists to weigh an investment’s opportunity against its risk. Yet, early accounts suggest readers are spending most of their time on one of those columns. Of its 408 pages, no section is given more ink than risk. Readers of the S-1, and of nearly all AI commentary, risk neglecting the work that is both hardest and most necessary: assessing AI’s risks and opportunities together rather than one at a time.

As much as the AI realm has gained a reputation for drama and politicking, this is a seriously weighty document, not a blog post or a high-minded essay. It serves as the basis for billion-dollar decisions. It’s a regulatory obligation, made with the intent to provide investors with material, accurate, and transparent information to evaluate the risks and opportunities associated with an investment.

The science of analyzing AI opportunity and risk side-by-side is a rare science frequently undercut by our human bias toward spectral extremes. However, it is a science that both investors and society writ-large must embrace not only in evaluating AI companies, but also in understanding the broad capabilities that flow from their data centers.

The Single Column Problem: Why AI Coverage Reports Risk or Opportunity, Not Both

The public debate tends to report one column at a time: risks or opportunities. Rare are those discourses that afford room to both. Even a Pollyanna will acknowledge that extremes make for better headlines. Further complicating matters: the trillions of dollars sitting piled atop the proverbial table mean that no party is disinterested.

Lest we fall prey to the folly of recency bias, we should note that new technologies never come with just a single column. Transformational technologies transcend binary “good” or “bad” classification; you needn’t strain for historical examples.

• The printing press mass-produced the Malleus Maleficarum and fuelled a century of religious wars.
  • It also spread scientific knowledge and literacy to the masses.
• The Haber-Bosch process provided World War I Germany with nitrates for the production of explosives.
  • It also enabled synthetic fertilizer and fed half the world’s population.
• The Internet distributes hate speech, scaled scams, and misleading information.
  • It also connects more than six billion people to the sum of human knowledge, and to each other.

But there’s an obvious difference between this moment and those historical examples: time has filled in those ledgers. As Steve Jobs famously said, “You can only connect [the dots] looking backwards”. Gutenberg and his contemporaries couldn’t have known that the same machine would spread witch-hunting manuals and religious wars while simultaneously building mass literacy and fuelling the Scientific Revolution. The early Internet was sold as connection and knowledge. Scaled scams and disinformation showed up on the ledger much later.

At our torrid pace, no one can tally every entry on AI’s ledger. This piece will not attempt to do so. The more useful exercise is to understand how the two columns, risks and opportunities, relate to each other. That analysis reveals a critical truth: nearly every entry comes with a twin in the other.

Two Blossoms of the Same Branch: How AI Risks and Opportunities Are Correlated

That pairing is not a coincidence; AI opportunities and risks are highly correlated, blossoming from the same branch of capability. Thus, risks and opportunities grow in tandem.

Look at nearly any breakthrough and you can find the same pattern. Models that understand proteins can design vaccines. The same understanding can help someone build a lethal pathogen. Models that find software flaws enable defenders to patch them faster. They empower attackers to exploit them faster too. An AI agent that acts on its own can clear a day’s worth of admin work. It can also take potentially destructive actions of which you did not approve.

You cannot cut risk without cutting opportunity. That leaves us as gardeners of a single branch, and a gardener’s first instinct is to prune. Pruning does exorcise the risk, but removes the benefit as a consequence. Reconciling AI’s opportunity and risk means tending the branch so that the benefit outgrows the risk.

Two things make this possible. First, tending inside and outside of the model: screening, sandboxes, humans-in-the-loop, physical system constraints (more on that in a moment). Second, the capability that creates a risk is frequently the best tool for mollifying it.

Biology, one of the AI capabilities most frequently cited in P(doom) discourses (the probability that AI causes catastrophe) due to the potential scale of the associated disaster, presents what might be the most challenging of all the aforementioned risk-opportunity dyads. OpenAI, Anthropic, xAI, and Google DeepMind have all addressed the potential negatives of AI’s knack for biological research by integrating safeguards into their systems, some more robustly than others. In fact, in the course of researching for this article, Claude’s classifier knocked me down from Opus 5.5 to Opus 4.8 on three occasions. 

Switched to Opus 4.8 Edit and retry with Opus 5.5

Opus 5.5’s safeguards flagged this message. You may be seeing this for the first time on an Opus model: Opus 5.5 is more capable and has stronger safeguards as a result, which can sometimes flag biology-research-adjacent work. We’re improving these safeguards to reduce the amount of incorrectly flagged messages. Send feedback or learn more.

Details: [bio]

This is useful insofar as it generates friction for would-be nefarious actors, but it is insufficient in that it will do little to sidetrack any attacker with a modicum of perseverance. As O.H. Scharfman and Will Bellamy argue, open-weight models trail the frontier by less than a year and can be run privately, unencumbered by the restrictions imposed by frontier labs’ safeguards. Thus, the greatest measure of leverage sits outside of the models in physical chokepoints like KYC screening at DNA synthesis providers and early-warning systems that read wastewater and clinical samples for threats before they proliferate. AI makes both of these approaches possible and is bulwarked by the fact that defenders, in Bellamy and Scharfman’s words, “generally have more access to capital, research, infrastructure, and capable personnel”.

Laying this approach atop the single-column reading of risk reveals what is at stake. If we read only the risk, our response will be to kneecap the model. In doing so, we forfeit the biological advances associated with projects like AlphaFold (used by over 3 million researchers to shave years off of lab work) and AI-augmented vaccines that are already passing human trials. All while still failing to ward off attackers using open-weight models.

The response changes when one simultaneously assesses both columns. Instead of a weaker model, stronger defenses around a capable one. Instead of pruning the branch, invest more time and resources into tending it. This is a choice, and someone has to make it.

Finance as Stewards: How Investors Shape AI Development

Finance is in a unique position to shape that decision. Anthropic’s S-1 is a rarity because the law obliges it to be. Regulation compels issuers to set forth both sides of the ledger, risk and opportunities, in the same document, audited and signed. Almost nothing else in this AI saga comes bundled that way. Successes get press releases and glistening demo reels, failures get bolded headlines and hand-wringing. Each comes to us as a single column.

That leaves the hard work of reconciliation to the reader. Thankfully, in finance, the reader holds levers. By deciding where capital gets allocated, how shares are voted, and what companies disclose, investors can decide what happens to these twisted branches and their dueling blossoms.

Those earlier ledgers of breakthrough technology took centuries to fill. Amidst the rapid pace of AI and our immense access to information, we do not have that luxury, nor that excuse. Today’s allocations, votes, and activism are tomorrow’s history. Investors cannot wait for the dots to connect. They must place them.

The Reconciliation Mandate: Assessing AI Risk and Opportunity Together

We cannot yet scrutinize AI’s completed ledger. Barring AI’s acceleration of longevity science, we will likely not have this opportunity in our lifetimes. But history tells us how such ledgers should be read. The printing press brought the scientific revolution and the witch trials; Haber-Bosch fed the world and armed the Kaiser. Our hindsight strikes neither column from the record. It teaches us to hold both at the same time.

AI will not wait for hindsight. So the discipline that Anthropic’s S-1 imposes by law, we must impose by habit. That is the net assessment. It is not a verdict we deliver once the dust settles but rather a balance we strike while it is still in the air.  Watch and respond to the risks, but tend to what grows beside them.