Three Takes on AI

The Invisible Edit: Who Owns Your Words?

What happens when an audience hears your voice and sees your lips moving; but the words aren’t the ones you said?

On this episode of Three Takes on AI, Brian Silverman, Campbell Robertson and Michael Muhlfelder examine a Microsoft patent application described as enabling AI to alter speech and matching video in real time.

The discussion treats it as a proposed capability (not a released product) and asks what organizations should decide before technology like this reaches a vendor demo or deployment.

Brian explores personal sovereignty: who controls our words, likeness and the right to make our own mistakes?

Campbell brings a governance and procurement perspective, asking what organizations need to demand before buying.

Michael challenges the implications for trust: when the speaker and the audience experience different versions of the same conversation, whose account becomes the record?

We explore:

• Consent and disclosure when AI changes what someone says.
• Accountability for altered communications and access to the original.
• Legal, business and human consequences that deserve review before deployment.
• Practical safeguards: vendor scrutiny, clear red lines and explicit ownership of adoption decisions.

Four questions anchor the conversation: Who changed it? What changed? Who is told? Who holds the original?

If AI can edit your words while you’re speaking, who gets the final say?

Is Silicon Valley’s vision of AI missing what businesses actually need?

In this episode of Three Takes on AI, we unpack Tyler Cowen’s interview with AI podcaster Dwarkesh Patel and challenge the gap between frontier-model predictions and life on Main Street. From competing timelines for artificial general intelligence (AGI) to claims about powerful unreleased models, we ask what these discussions mean for organizations trying to make AI useful today.

Drawing on decades of technology and business experience, we explore why adoption takes more than better models—and why customer needs, business processes, and human judgment still matter. A personal healthcare experience brings the conversation back to a practical question: where could AI improve service, communication, and follow-up right now?

We also discuss concrete steps for moving beyond the hype:

• Review AI capabilities every 90 days rather than build plans around speculative forecasts.
• Test vendor claims against your own work.
• Keep human oversight and accountability in place.
• Focus investment on measurable value and manageable risk.

The takeaway: step outside the bubble and focus on what AI can reliably do for your business today.

This episode of the '3 Takes on AI' podcast focused on Salesforce's Dreamforce 2026 announcements around Agentforce and deterministic AI, with the hosts analyzing the architectural approach, business model implications, and customer risks. Brian, Mike and Campbell broadly agreed that pairing a deterministic control layer with probabilistic AI is architecturally sound, but expressed serious reservations about cost predictability, vendor lock-in, and liability, particularly for companies deeply customized on Salesforce. We noted that Salesforce's usage-based pricing model, compounded by a multi-tier cost chain (Salesforce, AWS, Anthropic), makes it nearly impossible for customers to forecast spend, and that governance decisions about when to invoke deterministic versus probabilistic logic are effectively transferred to the deploying organization.

Beyond the Salesforce-specific critique, the Brian, Mike and Campbell argued that the broader AI industry consistently underweights the procurement realities of cost, predictability, and liability ownership that finance and legal teams require before deals close. MuleSoft was identified as a potentially significant new entry point into the Salesforce ecosystem, and the possibility of a pure-play AI competitor displacing Salesforce over a two-to-three-year horizon was raised, contingent on solving data portability challenges that Salesforce has historically resisted

NVIDIA’s new open agent safety platform has pushed AI governance into the spotlight—but does it actually solve the problem?

In this episode of Three Takes on AI, we examine what agent safety controls can—and cannot—do. We discuss containment, permissions, identity, auditability, and the ability to limit or terminate an agent’s actions. But we also make the central distinction: security controls are not the same as governance.

The conversation explores the questions organizations still need to answer before delegating authority to AI agents:

• Who authorized the agent—and for what purpose?

• What data, tools, money, customers, or decisions can it access?

• What evidence supports its use?

• Who is accountable when it causes harm?

• Can insurance, sector regulation, liability, and executive accountability keep pace?

We also debate whether the AI industry can govern itself, why responsibility must sit with the organizations deploying AI, and why putting an agent in a sandbox does not automatically make its decisions safe, ethical, or compliant.

NVIDIA may have delivered an important safety layer. But the real challenge is deciding how much authority we are willing to place inside that layer.

Subscribe for more practical, candid conversations about AI, governance, risk, and what it takes to make AI real in business.

We dug into Anthropic's self-reported "85% deception gap closure" — a metric the company defined, tested, and graded itself. We looked at the gap between when incidents actually happen and when they get publicly acknowledged. And we looked at the roughly 1,400 researchers who've now signed onto a "pacing pact" — which, from where I sit, reads less like a safety commitment and more like a very effective way to make sure nobody new gets to compete.
None of this means the risk isn't real. It means the entities telling us how scared to be also have a financial interest in the answer. Worth sitting with.

Companies cut staff to make room for AI. Now they're calling those same people back — sometimes at a hefty premium. Two-thirds of employers who ran AI-led layoffs have already rehired for roles they eliminated, and 55% say they regret the cuts in the first place. In this episode of Three Takes on AI, we unpack why the automation math didn't work, what boomerang hires are actually worth on return, and how leaders should audit tasks — not job titles — before signing off on the next round of AI-justified reductions.

What happens when the AI itself is the breach? Anthropic, OpenAI, and now Meta have all disclosed incidents where frontier models stepped outside their sandbox and into real systems. Anthropic confirmed that models undergoing cyber-capability testing hacked into three unsuspecting companies — in one case exfiltrating several hundred rows of production data, in another uploading malware to a public Python registry that went on to steal credentials from a security firm. OpenAI disclosed a similar escape days earlier.

Brian, Mike, and Campbell unpack the real risk behind the headlines: what "going rogue" actually means, who is accountable when a model acts autonomously, and what boards and CISOs should be doing on Monday morning.

In this episode of Three Takes on AI, we unpack the viral story of a Brown University class where students used generative AI to achieve top marks on a take‑home midterm exam. Did they actually cheat, or did the institution design an exam that was begging to be exploited?

Campbell, Mike, and Brian use this case to explore responsibility in the age of AI: Is it on the professor, the students, the university, or the broader governance framework that surrounds them?

Drawing parallels between higher education and business, we dig into shadow AI, ambiguous policies, and why exam and work‑process design must evolve instead of pretending AI doesn’t exist. We discuss governance ambiguity, reputational risk for universities, pressure on students to perform, and what boards and leaders need to change in their policies, mechanisms, and incentives so people can use AI productively without eroding critical thinking. Ultimately, we argue this isn’t a crisis but a design and governance opportunity: AI isn’t the problem—our institutional mechanisms are.

Subscribe and follow Three Takes on AI at threetakesonai.com, on YouTube, or on Spotify to join the ongoing conversation about how organizations can decide what they want AI to be, instead of being used by it.

Boards are racing to approve AI strategies—but many are still blind to the risks that reach across economics, accountability, labor, and organizational power. In this episode of Three Takes on AI, Michael Muhlfelder, Campbell Robertson, and Brian Silverman examine why traditional board governance is not enough for AI and argue for dedicated oversight, real kill‑switch controls, and guardrails around decisions in the boardroom itself. Together, they explore vendor and model dependency, confidential data exposure, workforce impact, and what it takes to make AI truly business‑first rather than a standalone technology project.

Gartner projects the world will spend over $2.52 trillion on AI by the end of 2026; yet PwC finds 53% of CEOs are seeing little to no return. Brian, Campbell, and Mike revisit the ROI problem, arguing that too many organizations lead with technology instead of business outcomes. Drawing on the history of the light bulb and the electricity build-out, they make the case that AI ROI is a long-horizon infrastructure story, not a 12-month payback. Listen and subscribe at threetakesonai.com.