Artificial General Intelligence: The Most Consequential Undefined Term in the World, and OpenAI's Inconsistent Take on It.
The Enlightened Spectator | Spectator Report
Executive Summary
In December 2015, OpenAI opened its doors with a mission statement that incorporated this concept: artificial general intelligence.
Ten years later, OpenAI’s CEO Sam Altman called AGI “a very sloppy term.” By summer 2025, he had upgraded the characterization: “not a super-useful term.”
That trajectory is the story here.
The term “AGI” has anchored billions of dollars of fundraising. It is in the background of a corporate governance crisis that nearly crippled OpenAI. It is featured in a legal contract between OpenAI and Microsoft that restructures their partnership upon its achievement. It is discussed in congressional testimony. And it has shaped a decade of coverage positioning OpenAI as humanity’s most consequential institution.
But AGI has had at least four definitions along the way. And none of them are philosophically coherent.
This report tracks the definitional arc of AGI from OpenAI’s founding documents through Altman’s public retreat. And it highlights the detectable signatures of Strategic Ambiguity.
This Report also includes a philosophical analysis that examines what ten years of definitional instability reveals about the quality of thinking underlying the most consequential technology debate of our time.
The Record: Only facts passing scrutiny under the Spectator Rules of Evidence.
Excluded Evidence: Reported claims failing the evidentiary rigors of the Spectator Rules of Evidence.
Perspectives: Major viewpoints disclosed without tribal signals or partisan labels.
PSYOP Report: Assessment of influence patterns in the coverage.
The ANALYSIS: Philosophical and epistemological examination of what the record reveals.
I. The Record.
The Record contains only facts that pass the evidentiary standards of the Spectator Rules of Evidence. The language is neutral. Characterization is avoided.
The Founding Definition
OpenAI was incorporated on December 11, 2015, as a nonprofit organization. Its stated mission was to ensure that artificial general intelligence benefits all of humanity.
OpenAI published its Charter in April 2018. The Charter states: “OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.” This is the publication’s founding, governing, and publicly filed definition of AGI.
The Charter further states that OpenAI “will also consider our mission fulfilled if our work aids others to achieve this outcome.”
The Commercial Transition and Its Contractual Consequences
In 2019, OpenAI created a for-profit subsidiary, OpenAI LP, under a capped-profit structure. The nonprofit retained governance control.
As part of OpenAI’s partnership agreements with Microsoft, a clause was inserted providing that Microsoft would lose commercial access to OpenAI technology upon OpenAI’s achievement of AGI. OpenAI’s public website states that such a system is “explicitly carved out of all commercial and IP licensing agreements.” OpenAI’s nonprofit board holds the authority to determine when AGI has been achieved.
Microsoft has invested at least $13 billion in OpenAI under agreements incorporating this clause.
The Information reported in 2023, citing documents it had obtained, that a Microsoft-OpenAI agreement defined AGI as a system capable of generating $100 billion in profits. Business Insider subsequently reported on The Information’s reporting. The original agreement is not publicly available. [See Excluded Evidence, Item 2.]
The Public Redefinitions
In February 2023, Altman published “Planning for AGI and Beyond” on the OpenAI website. The post states: “Our mission is to ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity.” The phrase “generally smarter than humans” differs from the Charter’s “outperform humans at most economically valuable work.”
In July 2024, OpenAI shared an internal five-level progression framework with Bloomberg, which Axios reported. The five levels are: Chatbots (AI with conversational language), Reasoners (human-level problem-solving), Agents (systems that can take actions), Innovators (AI that can aid in invention), and Organizations (AI that can do the work of an organization). The framework does not identify AGI as a specific level within the progression.
Multiple outlets reported that in a Bloomberg interview published January 5, 2025, Altman described AGI as having “become a very sloppy term” and stated, “If you look at our levels, our five levels, you can find people that would call each of those AGI.” Time and PYMNTS both cited this Bloomberg interview. [See Excluded Evidence, Item 3.]
On January 6, 2025, Altman published “Reflections” on his personal blog. The post states, “We are now confident we know how to build AGI as we have traditionally understood it.” The post does not specify which prior definition “as we have traditionally understood it” refers to. The same post states that OpenAI’s attention is turning to superintelligence, which it describes as “future AI systems dramatically more capable than even AGI.”
The Governance Crisis and Definitional Stakes
On November 17, 2023, OpenAI’s board voted to remove Altman as CEO, stating he had not been consistently candid in his communications. He was reinstated on November 22, 2023, following pressure from employees and investors. Microsoft, which had not been notified in advance of the board’s decision, was given a non-voting observer seat on the reconstituted board.
The board’s authority to declare AGI achieved—and thereby trigger changes to the Microsoft commercial relationship—resided with whatever board was in control of OpenAI at any given moment. This structural feature was present throughout the governance crisis.
The Retreat
Fortune reported in August 2025 that Altman had called AGI “not a super-useful term” at a CNBC appearance that summer. The same report noted that Eric Schmidt, who in April 2025 had told a conference that AGI would arrive within three to five years, had by August urged Silicon Valley to stop fixating on superhuman AI. Fortune reported that Max Tegmark, president of the Future of Life Institute, stated that calling AGI “not a useful term” was “a way for the company to steer clear of regulation while continuing to build toward more and more powerful models.”
On October 28, 2025, Microsoft and OpenAI announced a restructured partnership. OpenAI’s website noted the update without publishing the full terms of the revised AGI clause provisions.
II. Excluded Evidence.
Claim 1: Sam Altman posted on X in 2024 that OpenAI had “AGI achieved internally.” Fortune’s August 2025 report documented this post as part of its coverage of the definitional vibe shift. Altman subsequently posted on X in response to online speculation about internal AGI achievement: “Twitter hype is out of control again.”
Ruling: EXCLUDED from The Record as an established factual claim about AGI’s status. Admitted as a documented public statement by a named official through a verified institutional account.
The statement is authenticated under SRE Rule 902(f) as to authorship. It is admitted as proof the statement was made, not as proof that OpenAI achieved AGI. Altman’s subsequent X post characterizing the surrounding online speculation as “out of control” materially complicates the evidentiary weight of the original statement. A factual claim made and publicly walked back by the same named official (through the same platform, in the same evidentiary period) does not establish the underlying factual assertion. The statement belongs in The Record only as a documented official communication.
Standard: SRE Rule 902(f): self-authentication as to authorship, not content accuracy. SRE Rule 801(d): admitted as a documented statement, not as proof of the matter asserted.
Claim 2: The Microsoft-OpenAI 2023 agreement defined AGI as a system capable of generating $100 billion in profits. This claim has been widely repeated in subsequent coverage as the operative contractual definition of AGI.
Ruling: CONDITIONALLY ADMITTED at reduced evidentiary weight. The original Microsoft-OpenAI agreement is not a publicly available document and was not directly accessed. This publication’s treatment of its content rests on The Information’s reporting of the agreement’s terms, which Business Insider subsequently reported. This is secondary evidence of document content under SRE Rule 1002.
The Information is a Tier A outlet with a demonstrated record of obtaining and accurately characterizing private corporate documents. Its access to the underlying documents is described in its reporting. The claim has not been contradicted by Microsoft or OpenAI, and its substance is consistent with the AGI clause structure described in other documented public sources. On this basis, the claim is conditionally admitted in The Record as documented reporting from a Tier A outlet with described access to the original, but not as direct evidence of the agreement’s text. The original agreement was not accessed. The best evidence has not been disclosed.
Standard: SRE Rule 1002(b): direct reliance required where document content is material. SRE Rule 1004(c): secondary evidence conditionally admissible where Tier A outlet had demonstrated access; tiered and disclosed accordingly.
Claim 3: Multiple outlets reporting on Altman’s January 5, 2025 Bloomberg interview present the “sloppy term” characterization and the five-levels statement as independently corroborated facts established across multiple sources.
Ruling: ADMITTED with sourcing limitation disclosed. Time, PYMNTS, and LessWrong each cited these characterizations. All cite the same Bloomberg interview as origin. Multiple outlets reporting from the same single source does not constitute independent corroboration under SRE Rule 615. The underlying Bloomberg interview is paywalled and was not directly accessed. These statements are admitted as characterizations of Altman’s documented public remarks, attributed to multiple outlets reporting from a single Bloomberg interview, but not presented as independently corroborated.
Standard: SRE Rule 615: circular corroboration. Attribution language in The Record discloses the single-origin character.
III. Perspectives.
Significant viewpoints are described here without tribal signaling or political labels.
Perspective #1: The Definitional Evolution View.
The evolution of OpenAI’s AGI definitions reflects honest intellectual progress in a rapidly developing field. The Charter’s 2018 threshold—“outperform humans at most economically valuable work”—was a reasonable starting approximation for a field that didn’t yet know what it was building. The five-level framework represented a more architecturally sophisticated understanding of what general intelligence actually requires across dimensions of autonomy, reasoning, and novel action. Altman’s acknowledgment that AGI has become “a sloppy term” is scientific honesty—an admission that the field’s shared vocabulary has not kept pace with its technical progress. Every hard conceptual domain has this problem. Definitions in quantum mechanics, evolutionary biology, and consciousness studies have all shifted substantially as understanding deepened. Holding OpenAI to definitional precision that the broader scientific community hasn’t achieved sets a double standard. The goalpost moves because the original goalposts were rough approximations established at the beginning of the work, not because the company is evading accountability.
Perspective #2: The Strategic Ambiguity View.
The definitional instability is not principally a byproduct of honest intellectual progress. The definition has tracked operator advantage: expanding toward AGI proximity when capital was being raised and public positioning required scale, retreating when the term’s precision became a legal or regulatory liability. The board’s unilateral authority to declare AGI achieved—combined with a definition the board controls—functions as a financial instrument whose trigger the instrument’s holder defines. The retreat from the term in summer 2025, timed to increased regulatory scrutiny and the restructuring of the Microsoft agreement, is consistent with Max Tegmark’s regulatory-evasion characterization. If the definitional imprecision were purely epistemic, the company’s incentive would be to resolve it as quickly as possible, since precision would support rather than undermine its fundraising and public standing. Instead, the imprecision has been maintained across every context where its resolution would create accountability. That pattern is the finding.
IV. PSYOP Report.
Assessment of influence patterns in the coverage. Detectable influence signatures do not prove an intent to manipulate or influence; they can be cause knowingly or unknowingly.
Primary Tactic: Strategic Ambiguity
The influence pattern of strategic ambiguity is present in this media environment concerning AGI.
Strategic Ambiguity is maintaining ambiguous positions so different audiences interpret the same communication as supporting their preferred reading. Yet the operator retains maximum flexibility by avoiding binding commitments. So any subsequent action can be framed as consistent with prior statements regardless of what the action is.
Element 1: There are Multiple incompatible interpretations from the same source. From “generally smarter,” to the five-level framework where any level could be called AGI, to “not a useful term”—the arc tracks changes that aren’t clearly acknowledged.
Element 2: Operator retention of interpretive flexibility. OpenAI’s board holds unilateral authority to declare AGI achieved. No external party can make that determination, contest it, or compel a timeline.
Element 3: Different audiences resolving the ambiguity in their preferred way. Investors reading “we know how to build AGI” in January 2025 understood OpenAI’s foundational milestone as achievable and proximate. Regulators and journalists asking pointed questions were told the term was sloppy and not useful. Employees invoking the mission were told AGI is near and glorious. Each audience received the same institutional voice. Each resolved the ambiguity through their own prior framework, producing different apparent commitments from the same statement. The operator benefited from all resolutions simultaneously.
Element 4: Whichever interpretation is advantageous, the operator claims consistency. The January 2025 “know how to build AGI” declaration and the August 2025 “not a useful term” retreat both remain in the institutional record. Any position on AGI’s proximity—imminent, achieved, abandoned, irrelevant—can be framed as consistent with what OpenAI has previously stated. No statement was specific enough to disconfirm.
V. The ANALYSIS.
Here are a few philosophical points.
My argument is that OpenAI’s position on AGI is philosophically crude. (1) The term is poorly defined. (2) It ignores or conceals important metaphysical assumptions. And (3) it is unfalsifiable.
The Definition Problem.
OpenAI’s definition of AGI is inadequate. It doesn’t even try to satisfy criteria of a well-formed definition.
Suppose OpenAI chose to employ a philosopher to help them define the term. How might that philosopher approach the task?
A classical approach to definition is to identify a thing’s genus and specific difference. That means identifying the class or category a thing belongs to and then its differentia—it’s qualities that make it distinct from everything in the class. You would want the definition to include and exclude everything that it should.
You would then want to identify scope and boundary conditions. What clear cases fall within the scope of the definition? What clear cases fall without? And what cases would be borderline?
Next, you’d want to identify operational conditions. You’d want to specify how the thing could be identified, measured, or evaluated. This goes beyond what the term means to how you can know it when you see it. This is especially important in contexts where a concept must apply repeatedly and consistently. Minimally, you’d want to identify the properties, behaviors, or states that show the concept applies in a given instance.
And along the way, you’d do some metaphysical work to identify necessary and sufficient conditions. That is, you’d determine what must be present for the concept to apply, and what, if present, would be enough for the concept to apply. The metaphysician would also inquire into causation, ontological status, and essential versus accidental properties, among many other things.
For something like AGI, none of that is easy. And the problem compounds:
What even is intelligence?
What is general intelligence?
And what is artificial general intelligence?
I’m unsure that there is a robust consensus on even the first of those questions. It becomes less certain as you qualify it and then theoretically extend it to a computational process.
OpenAI’s “generally smarter than humans” does very little of the above. And it isn’t even clear what “smarter” means in this context.
I do not suggest OpenAI is intentionally manipulating anyone. I reserve judgment between the two perspectives above. The operational signatures of strategic ambiguity can be present in an information environment without an underlying intent to influence, manipulate, or deceive. It could just indicate genuine uncertainty as the company explores new territory.
My point is just that their definition presently is inadequate.
This may be stating the obvious.
But maybe not—significant business decisions are being made with real world consequences that presuppose an understanding of “AGI.” Yet I’m not sure anyone knows what that means.
Metaphysical assumptions.
OpenAI’s AGI definition assumes metaphysical commitments that are not apparent to non-philosophers. And they’re not apparent in the public messaging about AGI.
The definition makes assumptions about the nature of intelligence and whether computational systems can possess it. My best assessment presently is that it presupposes a view called “functionalism.”
Roughly, functionalism is a view that functional roles define mental states. So mental states like desires, beliefs, or pain are identified by their causal role in a system. Three interacting components determine them: inputs (stimuli), internal transitions (the causal interplay between the mental state and other internal states), and outputs (the resulting behavior).
A common implication of functionalism is the view that different physical substrates could realize mental states, i.e., not just human brains, but also silicone-based computer chips.
I presume OpenAI’s AGI position takes a functionalist view of intelligence. That being something like performing the right computational role within the system of inputs, outputs, and internal transitions.
But none of this universally accepted. I’ll give just one counter example.
In 1980, John Searle proposed the popular (in some circles) “Chinese Room thought experiment.” It addressed the nature of intelligence and whether computational systems can possess it.
The thought experiment was about like this: a person in a closed room receives Chinese characters through a slot. The person is monolingual and doesn’t know Chinese. They then apply rules from a huge rulebook to produce appropriate Chinese character responses. The rulebook explains how to manipulate the symbols based on their shapes. But the person inside the room doesn’t understand what any of the symbols or shapes mean. They mechanically apply the rules, however, to manipulate the symbols to produce responses, and they slide these outputs to the people outside the room. And the people outside the room believe they’re communicating with a Chinese speaker inside the room. The Chinese Room seems to “know” Chinese.
The thought experiment illustrates a system that produces intelligent-seeming outputs through the formal manipulation of symbols without an understanding of what the symbols mean.
Searle’s conclusion was that syntax is not sufficient for semantics: a system can produce something that resembles intelligence but lacks understanding, intentionality, or meaning.
This is a very thin sampling of many viewpoints and debates within the Philosophy of Mind. The point here is not to argue Searle’s argument is a perfect defeater of functionalism. Or even that functionalism is right or wrong.
The point here is just that the AI industry has apparently assumed these issues are resolved in favor of functionalism. They’re not.
It seems the industry presupposes contested philosophical commitments that they haven’t made clear:
Mental states reduce to functional role. That’s unresolved.
Intelligence is defined by what a system does rather than what it is or something else. That’s unresolved.
Behavioral equivalence is ontological equivalence. That’s unresolved.
These commitments have major implications. And they are in the background of discussions about machine consciousness, which is another can of worms.
But these types of presupposed views have apparently driven business strategy in a multi-billion-dollar industry. They are shaping public discourse on how we understand our own humanity. And they raise ethical questions I won’t explore here.
If functionalism is wrong, then the AGI discussion has assumed a category error from the start.
Falsifiability
Vague definitions risk being unfalsifiable.
In the philosophy of science, claims are considered scientifically meaningful when they provide conditions of falsification. These are predictions that evidence could contradict.
But if a definition is gerrymandered to be compatible with every possible outcome, it isn’t falsifiable.
Here’s why that matters: an unfalsifiable claim can be “confirmed” by anything but never really tested or refuted.
Unfalsifiable claims are scientifically meaningless or weak.
The AGI definition is vague. It is unclear what evidence would falsify hypotheses about whether it is achieved. And at worst, industry leaders can adjust the definition to avoid falsification whenever it is commercially beneficial to do so.
A definition of AGI would be more scientifically meaningful if it set clear predictions that could be tested and disproved by evidence.
* * *
OpenAI’s AGI position could benefit from more thoughtful analysis. Philosophically, it would be more sophisticated if it did this:
Crafted a clearer definition
Disclosed its metaphysical assumptions transparently,
And set conditions of falsification to test hypotheses about whether AGI is achieved in a given instance.
There may be incentives for not doing that. Or maybe they’re sincerely doing the best they can.
I mostly reserve judgment on many claims about AI. But I’m skeptical of “AGI.” And I think Altman’s acknowledgment that AGI is a “very sloppy term” is a slight understatement.
Source Map
Primary Documents and Official Sources
OpenAI Charter — “OpenAI’s Mission and Principles.” https://openai.com/charter/
OpenAI — “Our Structure” (updated October 28, 2025). https://openai.com/our-structure/
Sam Altman — “Planning for AGI and Beyond,” OpenAI (February 2023). https://openai.com/index/planning-for-agi-and-beyond/
Sam Altman — “Reflections,” personal blog (January 6, 2025). https://blog.samaltman.com/reflections
OpenAI — “The Truth Elon Left Out” (internal emails on for-profit transition). https://openai.com/index/the-truth-elon-left-out/
Contractual and Governance Record
Business Insider / AOL — “OpenAI and Microsoft Have Put a Price Tag on What It Means to Achieve AGI.” Reports on The Information’s reporting, which cited documents obtained directly. The Information original is paywalled and was not accessed by this publication. https://www.aol.com/openai-microsoft-put-price-tag-050016426.html
OpenAI — “Our Structure” (AGI clause and Microsoft commercial terms disclosed). https://openai.com/our-structure/
CNBC — “Former OpenAI Board Member Explains Why CEO Sam Altman Was Fired” (May 2024). https://www.cnbc.com/2024/05/29/former-openai-board-member-explains-why-ceo-sam-altman-was-fired.html
Time — “An OpenAI Timeline: Musk, Altman, and the For-Profit Shift.” https://time.com/7328674/openai-chatgpt-sam-altman-elon-musk-timeline/
Documented Public Statements
Time — “How OpenAI’s Sam Altman Is Thinking About AGI and Superintelligence in 2025” (reporting on Bloomberg interview, January 2025). https://time.com/7205596/sam-altman-superintelligence-agi/
PYMNTS — “Sam Altman: OpenAI’s New Model Passes AGI Threshold” (reporting on Bloomberg interview, January 2025). https://www.pymnts.com/artificial-intelligence-2/2025/sam-altman-openais-new-model-passes-agi-threshold/
Fortune — “AGI Talk Is Out in Silicon Valley’s Latest Vibe Shift” (Sharon Goldman, August 25, 2025). https://fortune.com/2025/08/25/tech-agi-hype-vibe-shift-superpowered-ai
Capability Framework Documentation
Axios — “OpenAI Nears ‘Reasoning’-Capable AI” (five-level framework, July 2024). https://axios.com/2024/07/15/openai-chatgpt-reasoning-ai-levels
LessWrong — “We Know How to Build AGI — Sam Altman” (Bloomberg interview context and sourcing). https://www.lesswrong.com/posts/T5p9NEAyrHedC2znD/we-know-how-to-build-agi-sam-altman
Kaizenko — “OpenAI’s Five Levels of AI: A Roadmap from Chatbots to AGI.” https://www.kaizenko.com/openais-five-levels-of-ai-a-roadmap-from-chatbots-to-agi/
Coverage and Critical Analysis
Machine Intelligence Research Institute — “Comments on OpenAI’s ‘Planning for AGI and Beyond’“ (Nate Soares, March 2023). https://intelligence.org/2023/03/14/comments-on-openais-planning-for-agi-and-beyond/
Time — “OpenAI Wants to Go For-Profit. Experts Say Regulators Should Step In” (November 2025). https://time.com/7279977/openai-for-profit-letter-elon-musk/
Philosophy of Mind
Stanford Encyclopedia of Philosophy — “The Chinese Room Argument.” https://plato.stanford.edu/entries/chinese-room/
Encyclopaedia Britannica — “Chinese Room Argument.” https://www.britannica.com/topic/Chinese-room-argument