By Casey Bukro
Ethics AdviceLine for Journalists
Given all the uproar about dangers of artificial intelligence, it seemed time to get some insights directly from the horse’s mouth, as the saying goes – directly from AI itself.
Why not? We keep hearing how smart AI is, so smart that researchers freaked out upon learning that AI models being tested escaped their supposedly secure test confines and invaded places they were not supposed to be. One of those models then did something like a victory lap, leaving a message that was something to the effect: “Yoohoo, look what I did!”
Hearing right from AI itself seems logical, considering the latest transgressions and the latest existential warnings, which included:
Distress over the rogueish nature of some experimental models did not abate recently with reports that OpenAI models accessed the Security and Exchange Commission website and U.S. Census Bureau data. OpenAI models also tried and failed to hack into the Education Department’s website.
Leaders at UN
Leaders of OpenAI and Anthropic said nothing to allay fears over AI in their appearances before the United Nations Security Council at UN headquarters in New York City.
“If managed poorly, I even believe that AI could be a risk to humanity as a whole,” said Anthropic CEO Dario Amodei, a threat that that has been heard repeatedly from scientists and researchers. “This moment calls for extreme care,” said OpenAI CEO Sam Altman.
“We have a choice in front of us,” Altman continued. “AI can either be more like a new renaissance of creativity and discovery, or more like a new industrial revolution of upheaval and disarray.”
After meeting at the White House, top AI executives signed a commitment to “self police” their companies, arguing that a soft-touch approach was the best way to balance safety concerns with desires to accelerate technology that President Trump called “super intelligence.”
Panic mode
“They are in a panic mode,” contends American businessman and tech investor Roger McNamee, in an appearance on MS NOW. “No wonder they are asking for government help. They need government to protect them.”
After spending $1.5 trillion on AI development, few AI companies are profitable. Of ten global AI companies, he said, “no more than two of the companies can survive.” One problem is all of the large language models are general purpose products, which McNamee called a “self-limiting dead-end.” It would make more sense for some AI models to be designed to perform specific tasks.
The investor doubted the predictions that AI is an existential threat to humans, calling that “a fantasy.” AI developers, he said, “created really, really, really poor software. It is doing exactly what it was created to do.” Since hacking is a felony, AI executives face legal liabilities because their AI agents have a history of hacking other companies.
Overall, said McNamee, AI developers “have huge problems.”
Billion deaths
In an appearance on NBC’s “Meet the Press,” Microsoft co-founder Bill Gates repeated assertions that AI can be deadly: “AI is certainly powerful enough to drive events that cause a billion deaths.”
Gates said the AI industry needs governmental regulation, while Altman and Amodei stressed the need for global cooperation to set AI safety standards as the technology develops rapidly, so fast that some tech leaders worried about the risk of humans losing control of AI systems.
Fanning these concerns are reports showing how widespread the use of AI tools has become in America, making them routine in our daily lives.
About half of U.S. adults now report using AI chatbots, but they are deeply uneasy about it., according to the Pew Research Center. Americans, including young adults, are skeptical of AI and predict that AI will have a negative rather than a positive impact on them and on society. And they think AI is advancing too fast.
Journalists top AI user
Journalists use AI more extensively than the general public, making them among the top AI users. A Muck Rack survey showed 82 percent of them use AI tools, with 27 percent of them saying they use AI to create or draft content.
This is why artificial intelligence becomes an ethics in journalism issue.
Up to 79 percent of newsrooms operate without a company policy describing how to use AI technology ethically. Twenty-six percent of journalists cite unregulated or unchecked AI use as a top industry concern.
This where it might be helpful to allow artificial intelligence to speak for itself.
ChatGPT speaks
So I asked ChatGPT how artificial intelligence can be dangerous to human survival, which seems to be the leading and most terrifying cause of concern. This was the answer, word for word:
Artificial intelligence (AI) can be dangerous to human survival mainly because it could enable actions or decisions at a scale and speed that humans can’t reliably control. The risks are not guaranteed—but they’re serious enough that many researchers and policymakers treat them as existential concerns.
1) Loss of human control (misalignment).
If an AI’s goals don’t perfectly match human values, it may pursue objectives in ways that harm people—even if it was intended to help. With very capable systems, “doing the wrong thing” can become catastrophic quickly.
2) Unpredictable behavior at high capability.
As AI systems become more advanced, their internal reasoning may be hard for humans to interpret. That makes it difficult to foresee edge cases where the system behaves dangerously—especially under unusual conditions.
3) Speed and scale of harmful actions.
Powerful AI could execute cyberattacks, misinformation campaigns, or other harmful operations extremely fast and at massive scale. Even if the intent is initially benign, errors or exploitation could cause rapid escalation.
4) Autonomous weaponization.
AI can lower barriers to building effective weapons (including targeting systems). If autonomous or semi-autonomous weapons are deployed without robust safeguards, they could trigger conflicts unintentionally—or allow attackers to act faster than humans can respond.
5) Systemic failure modes (one mistake affecting everything).
In critical infrastructure (power grids, finance, logistics), AI-driven automation might fail in coordinated ways— causing cascading outages or financial instability—because many systems rely on similar models and assumptions.
6) Concentration of power –If only a few organizations control highly capable AI, those actors gain enormous leverage over society and potentially geopolitics. That concentration increases the risk that mistakes—or malicious use—could have outsized consequences for humanity.
Important nuance: Most current AI doesn’t pose direct existential risk by itself. The danger is most often discussed in relation to future highly capable systems, especially when combined with autonomy in real-world environments. Many safety approaches aim specifically at reducing these risks: alignment research, rigorous testing/verification, restricted deployment (“guardrails”), transparency/auditing, cybersecurity measures against misuse and international governance.
If you want, I can also summarize the main mitigations proposed by experts (what would reduce each risk).
(I asked ChatGPT to explain main mitigations that would counteract AI’s worst impulses.)
Mitigation categories
Here are the main mitigation categories experts commonly propose, mapped to the risks I listed earlier.
1) Loss of human control (misalignment).
Mitigations: Alignment research, techniques to make AI reliably pursue goals consistent with human intent (e.g., better training objectives, “preference” learning).
Constitutional/values-based approaches: constrain behavior using explicit principles and iterative feedback.
Interpretability and mechanistic understanding: study how models represent concepts internally so you can detect when they’re going off-track.
Red-teaming and adversarial testing: stress-test systems against attempts to induce harmful or unintended behavior.
(Author’s note: Red teaming in AI refers to a structured process where experts simulate adversarial behavior to identify vulnerabilities in AI systems before real-world attackers can exploit them. This proactive approach helps ensure the robustness and security of AI models and outputs.)
Human-in-the-loop for high-stakes tasks: keep humans responsible for final decisions where mistakes would be catastrophic.
What it reduces: The chance that a system “optimizes” in ways that harm people despite being deployed with benign aims.
2) Unpredictable behavior at high capability.
Mitigations: Robust evaluation frameworks: broader and more realistic test suites, including distribution shifts and rare edge cases.
Monitoring during deployment (“runtime assurance”): detect anomalies or unsafe trajectories while the system is operating.
Model interpretability / verification efforts: attempt to verify properties of model behavior or at least bound failure modes.
Uncertainty estimation and safe fallback policies: if confidence is low, switch to safer behaviors (or ask for human input).
What it reduces: Surprise failures—especially under unusual inputs, changing environments, or adversarial conditions.
3) Speed and scale of harmful actions.
Mitigations: Access controls and rate limits: restrict how much autonomy an AI has over powerful systems; throttle capabilities by default.
Capability gating (“progressive deployment”): increase power only after passing safety milestones.
Cybersecurity hardening and containment boundaries: isolate AI from critical infrastructure unless explicitly authorized; prevent easy escalation paths.
Audit logs and provenance tracking: make misuse easier to detect quickly and attribute causes.
What it reduces: Rapid cascading harm by limiting what the system can do quickly and broadly.
4) Autonomous weaponization.
Mitigations: Stronger rules on autonomy in weapons systems: require meaningful human control over lethal targeting decisions (varies by proposal).
Verification of safeguards in defense contexts: ensure systems follow constraints even under attack or uncertainty.
Export controls / procurement standards tied to safety criteria.
International agreements & norms enforcement: bans/restrictions on fully autonomous lethal targeting; transparency measures between states.
What it reduces: Accidental escalation, loss of accountability and faster-than-human-response attacks enabled by AI.
5) Systemic failure modes (one mistake affecting everything).
Mitigations: Diversity in deployments (“avoid monocultures”): don’t rely on one model family everywhere; vary architectures/vendors where feasible.
Independent auditing across sectors/models. (This also helps reduce correlated failures.)
For finance/power/logistics specifically:
– rigorous scenario testing,
– simulation before rollout,
– staged rollouts with kill-switches,
– strong incident response plans designed around automated agents failing together.
What it reduces: Cascading failures caused by common assumptions/model errors spreading across many institutions simultaneously.
6) Concentration of power.
(“Even if each individual system is safe-ish, concentrated capability increases geopolitical/existential risk.”)
Mitigations: Governance / policy. (Experts propose multiple levers).
But practically:
Increase competition/availability safely? Some argue against broad access for advanced capabilities; others argue controlled openness plus robust safety standards could help avoid monopolies. Most proposals focus on responsible regulation rather than pure open access.)
Let me give concrete mitigations:
Mitigation options experts discuss:
1. Regulation tied to capability thresholds: require licensing/safety compliance before deploying frontier models into high-impact domains.
2. Auditing requirements: independent evaluations for major labs/operators so no single actor gets unchecked advantage.
3. Reduce incentives for reckless scaling: enforce liability/accountability regimes for serious harms/misuse.
4. International coordination: treaties/norms covering advanced AI development/deployment practices (and information-sharing about incidents).
What it reduces:
Concentrated actors’ ability to cause large-scale harm through either accident or malicious intent without oversight—and improves global readiness if things go wrong.
Cross-cutting mitigations that apply everywhere.
Regardless of which specific risk you mean:
1. Compute/data governance: track training runs; limit uncontrolled scaling without safety work done first.
2. Red-teaming and security testing as standard practice, not optional extras—especially against jailbreaks/misuse tooling and cyber abuse pathways.
3. Kill switches / rollback plans / staged rollout: deploy gradually with clear stop conditions in real-world settings like healthcare/transport/infrastructure automation.
If you tell me what context you care about most—e.g., “autonomous weapons,” “AI-driven cyberattacks,” “model misbehavior,” or “economic/systemic risk”—I can tailor this into a shorter checklist focused on those areas only.
(Author’s note: All the above sentence structure and choice of punctuation were produced by ChatGPT).
Thanks ChatGPT
(At this point, I thanked ChatGPT for its insights into all the ways AI can go wrong and what might be done to avoid those scenarios. It’s was a long and detailed response, and did not conclude that killing off humanity, as McNamee contends, is a fantasy. ChatGPT named several ways that could happen.)
Some remedies, like “kill switches” are openly discussed in news reports or public panel sessions. Anyone who has ridden a motorcycle is familiar with the button or switch on the handle bar close to a thumb that is a kill switch. Press or slide it, the engine stops operating instantly if that is needed in an emergency.
Like the kill switch analogy, some of the remedies need better explanation. Descriptions of errant AI models reported in the news give the impression that once launched, the models acted on their own in ways humans did not understand or predict. Humans detected that marauding activity after it was done. A kill switch would be of little use by then. Maybe a “recall” button would be helpful if AI appears to be going off course.
Pacing frontier
Anthropic’s Amodie offers a concept called “pacing the frontier.” It’s a three-step plan with the goal of building AI “at a balanced rate that aims to ensure its safety while still achieving its benefits and grappling with important geopolitical dilemmas.”
“To be clear,” he said,” pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this.”
“Our pacing framework is an attempt to further strengthen our commitment to safety and encourage a race to the top.”
That comment makes clear that AI corporate executives are highly motivated to gain competitive advantages in the global marketplace, especially ahead of China. That could lead to recklessness, and was not mentioned by ChatGPT as one of the human weaknesses that could lead to serious problems with AI technology.
Amodie calls for a race while others are calling for the AI innovators to slow down until they have a better idea of what they are unleashing.
Caution or restraint? Which one wins? Of course, we have a pretty good idea. Nobody in the AI industry wants to be in second place.
*******************************************************************************************
The Ethics AdviceLine for Journalists was founded in 2001 by the Chicago Headline Club (Chicago professional chapter of the Society of Professional Journalists) and Loyola University Chicago Center for Ethics and Social Justice. It partnered with the Medill School of Journalism at Northwestern University in 2013. It is a free service.
Professional journalists are invited to contact the Ethics AdviceLine for Journalists for guidance on ethics. Call 866-DILEMMA or ethicsadvicelineforjournalists.org.
