On the inflation of expertise, the architecture of a misunderstood technology, the gap it is creating — and why distributing knowledge is not idealism, but structural necessity.
This article begins with a terminological complaint — the term "AI Expert" — and uses it as an entry point into a larger set of questions. What does the technology actually consist of, and why does that structure matter for how we think about who understands it? What happens when a technology this consequential is built on internal logic that even its creators cannot fully audit? And when the gap between those who understand it and those who do not begins to translate into concentrated wealth and power, what are the stakes of leaving that gap unaddressed?
These questions are connected. The article follows that connection from the surface of a label to the floor of a civilizational argument.
Every technological wave produces its own vocabulary of self-promotion, and we are currently drowning in one of the more irritating examples: the AI Expert. The title has proliferated with the kind of enthusiasm typically reserved for multi-level marketing schemes and motivational speaking. It appears on LinkedIn profiles, conference speaker bios, and consulting proposals with remarkable consistency and almost no definitional rigor.
This is not a minor semantic complaint. Imprecise language produces imprecise thinking, and imprecise thinking — when applied to a technology that is restructuring labor markets, knowledge production, and institutional power — produces decisions with consequences that outlast the buzzword that enabled them.
So let us be precise.
01 — What "Expert" Actually Requires
Expertise, in any rigorous sense, refers to deep, specialized knowledge within a bounded domain. A cardiologist is an expert in the cardiovascular system, not in medicine as a general enterprise. A structural engineer is an expert in load-bearing systems, not in construction as an abstract concept. The meaning of expertise derives precisely from its specificity — the narrower and deeper the knowledge, the more the label earns its weight.
"Artificial Intelligence," by contrast, is not a bounded domain. It is a descriptor — a shorthand for a vast, multi-layered technological ecosystem that spans energy infrastructure, hardware engineering, advanced mathematics, cognitive science, ethics, policy, and application design. Claiming expertise in "AI" as a whole is roughly equivalent to claiming expertise in "science." The statement is not technically false. It is simply too large to mean anything.
To understand why, it helps to look at what the technology actually is — not as a product category, but as a system.
02 — The Architecture of AI: Four Layers, Each a World
What follows is a simplified framework — a first-order decomposition of what "AI" actually refers to as a technological system. Each layer contains considerable depth, its own sub-disciplines, and its own class of specialists. This is not an exhaustive taxonomy. It is a map sufficient to illustrate why the term "AI Expert" collapses under its own weight.
AI is, before anything else, an energy problem — and not in the way that other digital technologies are. A conventional Google search consumes approximately 0.3 watt-hours of electricity. A single query to an advanced generative AI model like ChatGPT consumed an estimated 2.9 watt-hours in 2024 — nearly ten times as much.[1] Multiply that by billions of daily queries and the scale becomes structural, not incidental.
Training large models amplifies this further. Training GPT-3 consumed an estimated 1.29 gigawatt-hours of electricity. GPT-4's training is estimated at over 50 GWh — roughly equivalent to 0.1% of New York City's annual electricity consumption in a single training run.[2] GPU-accelerated AI server energy usage grew from under 2 TWh globally in 2017 to over 40 TWh by 2023.[3]
The IEA projects that global data center electricity consumption will grow from 415 TWh in 2024 to approximately 945 TWh by 2030 — with AI identified as the primary driver, and AI-specific server consumption projected to grow at 30% annually.[4]
The physical substrate of AI consists of Data Centers, Graphics Processing Units (GPUs), Tensor Processing Units (TPUs), high-density Servers, and High-Speed Networks. Each component is a specialized engineering field in its own right.
GPUs — originally designed for rendering graphics — turned out to be exceptionally well-suited to the parallel matrix computations that underlie neural network training. Nvidia's H100 chip became the primary hardware bottleneck for AI capability globally. By 2025, hyperscalers — Google, Meta, Amazon — were collectively estimated to spend $364 billion on data center construction in that year alone.[5]
This is the cognitive layer — the Large Language Models (LLMs), vision models, and multimodal systems (GPT, Sonnet, Gemini, Llama, and their successors) that most people encounter as "AI." At its foundation lie Machine Learning Algorithms, Artificial Neural Networks (ANNs), and architectural innovations like the Transformer — introduced in the 2017 paper "Attention Is All You Need" — the underlying structure of virtually every major LLM currently deployed.
This is graduate-level mathematics, statistical learning theory, and systems engineering operating simultaneously at enormous scale. The number of people on the planet with genuine, research-level competence in this layer is, relative to the number claiming "AI expertise," embarrassingly small.
The surface layer — the products, tools, and interfaces that most people interact with daily. Text generation, image synthesis, audio transcription, code assistants, data analysis tools, recommendation engines. This layer has enormous practical value and is the most visible face of AI. It is also, by a considerable margin, the layer requiring the least technical depth to operate effectively.
Building effective products on top of AI infrastructure requires real skill — prompt engineering, fine-tuning, system design, UX judgment, domain expertise. These are legitimate capabilities. The point is simply that they are capabilities in the application layer, which sits on top of three layers of technical complexity that remain largely invisible to those operating within it.
The above framework is, explicitly, a simplification. Each layer contains its own taxonomy of specializations, and the boundaries between them are messier in practice than any diagram suggests. The point is not comprehensiveness. The point is that a system this architecturally complex does not yield to a single label — and yet we keep applying one.
03 — The Tool User Is Not the Expert
A useful distinction, frequently collapsed in practice: there is a meaningful difference between being proficient with a technology and being an expert in it.
Consider a dentist who uses AI-powered imaging diagnostics to analyze X-rays with greater precision than any unaided human eye. This person is genuinely skilled, genuinely valuable, and operating at a high level of professional competence. They are an expert in dentistry. The AI is an instrument in their hands — one they use exceptionally well. They are not, by any reasonable definition, an AI expert. The sophistication of the instrument does not transfer to the user as a credential in the instrument's underlying engineering.
Proficiency with AI tools — even advanced, creative, or highly productive proficiency — is a legitimate and increasingly valuable skill set. It should simply be named accurately: competence in the application layer, domain-specific AI literacy, tool proficiency. Calling it AI expertise inflates the individual claim while deflating the meaning of expertise in the technology itself.
Someone who subscribed to an AI platform, completed a trending online course, and can now produce competent outputs has, at most, acquired human intelligence about AI tools. The technology itself remains substantially unexamined. Most of us drive cars without understanding combustion. The problem arises when we start claiming to be automotive engineers — and then get hired as ones.
04 — We Built Something We Do Not Fully Understand
The terminology problem would be merely annoying if the underlying technology were itself well-understood by those building it. It is not.
"We don't really understand exactly how they do those things."
— Geoffrey Hinton, Nobel Laureate in Physics (2024), on the neural networks he helped build. 60 Minutes, CBS News, 2023What Hinton is describing is not vague philosophical discomfort. It is a specific and technically grounded problem. When engineers design a neural network, they define its architecture and training objective. What they do not define, and cannot fully predict, is what the network will learn as a result of that training process. During training, billions of numerical parameters — called weights — are adjusted iteratively across vast datasets. The adjustments are mathematically precise. But the representations that emerge are not specified by the engineers. They emerge.
This is the core of the black box problem. We can measure what the model outputs. We can observe how it behaves across different inputs. What we cannot easily do is look inside and understand why it produced a specific output. The field of Mechanistic Interpretability (MI) exists precisely to address this gap — attempting to reverse-engineer the internal logic of neural networks by uncovering what researchers at ACM have described as "human-understandable circuits, algorithms, and causal structures that drive model behavior."[6]
In critical domains — healthcare diagnostics, credit assessment, legal analysis, hiring — a model might produce the correct output for the wrong internal reasons: a pattern that works reliably on training data but fails in ways that are invisible until a high-stakes edge case surfaces. As researchers in the field note, without mechanistic understanding, engineers are "stuck reacting to symptoms rather than preventing failures."[7]
The phrase "AI Expert," in this context, takes on an additional layer of irony. We are awarding expertise certificates for a technology whose foundational layer — the model itself — is not fully understood by the Nobel laureates who built its intellectual foundations.
05 — The Gap: Not Just Knowledge, But Wealth
The conversation about AI and inequality has typically been framed around labor markets — which jobs will be displaced, which will be augmented, which will be created. This framing is real but insufficient. The more consequential gap that is forming is not between employed and unemployed, but between those who own the technology and those who are subject to it.
Empirical research across the US, EU, and Japan from 1995 to 2020 finds a statistically significant positive correlation between AI capital stock accumulation and wealth disparity — with the effect strongest in contexts where wealth is already concentrated.[8] A 2025 study in the European Economic Review found that medium and high-skilled workers experience compression in their income share under AI adoption — driven by wage compression rather than employment changes — while capital owners capture a disproportionate share of the productivity gains.[9]
Larry Fink, CEO of BlackRock, wrote in 2026 that AI risks creating "K-shaped outcomes" — a bifurcated economy where firms and investors with access to capital benefit from accelerated growth, while those less exposed to rising asset valuations stagnate. Since 1989, he noted, median US wages have lagged stock market returns by a factor of fifteen.[10]
The two camps on this question are genuinely divided, and the disagreement is worth representing accurately:
AI will compress the skill premium by automating tasks previously reserved for high-skill workers, enabling lower-skill individuals to perform more complex work and narrowing wage gaps over time. Bloom et al. (2024) model this as a medium-term equilibrium effect. David Autor (2024) argues AI could "rebuild the middle class" by allowing workers without advanced credentials to access capabilities previously gated behind years of specialization.
Every major technological breakthrough in modern history — the industrial revolution, electrification, the internet — concentrated gains among capital owners before (and often without ever) broadly distributing them. Cazzaniga et al. (2024) find that around 60% of jobs in advanced economies face high AI exposure — and that AI adoption could widen the digital divide and global income disparity. Historical wage data from the robotic automation era shows workers in automated firms lost approximately 9% of annual wage earnings over five years.
The honest assessment is that both scenarios contain valid mechanisms — and that which one prevails depends heavily on policy choices, governance structures, and the degree to which access to AI capability is genuinely distributed rather than concentrated. The historical base rate is not reassuring.
06 — Why Distribution Is Not Idealism — It Is Structural Necessity
The argument for distributing AI literacy and access across populations is sometimes framed as a humanist or egalitarian aspiration — a nice thing to strive for, situated somewhere between a policy preference and a moral sentiment. This framing understates the case considerably. The argument is better understood as a structural requirement of functioning democratic systems.
John Rawls, in A Theory of Justice (1971), argued that just institutions must ensure that the basic structure of society does not systematically disadvantage individuals based on circumstances beyond their control. Applied to the distribution of knowledge, this logic has been developed into what is now called epistemic justice: the idea that a systematic lack of opportunity to acquire knowledge one needs as an individual and a citizen is itself an injustice, not merely an inconvenience.[11]
Jürgen Habermas argued in Between Facts and Norms (1996) that democratic legitimacy depends on the quality of deliberation — and that deliberation requires participants who can engage with the substance of what is being decided. When critical infrastructure is a black box to most of the people it governs, the discursive conditions for legitimate democratic governance are structurally absent.[12]
Miranda Fricker's concept of epistemic injustice adds a further dimension: when some actors are systematically excluded from the production and validation of knowledge — in this case, knowledge about how AI systems work, who benefits from their deployment, and what risks they carry — they are harmed not just economically but epistemically. They become subjects of a system whose logic they cannot participate in shaping.[13]
These are not abstract philosophical positions. They describe a concrete structural condition that is forming right now. AI systems are currently embedded in hiring decisions, credit assessments, medical triage, content distribution, and educational tools. The entity that shapes which information reaches whom — and under what framing, at what priority, filtered by what criteria — is no longer solely human. In that context, treating AI as a black box that only specialists need to understand is not epistemic humility. It is abdication.
The knowledge gap and the wealth gap are not separate problems. They are the same problem operating at different layers of the same system. Breaking open the black box — developing genuine, layered understanding of what this technology is, how it operates, where its limits are, and who controls it — is a precondition for democratic agency in systems that now constitute daily life.
Imprecise language about AI is not a harmless quirk of a fast-moving industry. It is a mechanism by which real complexity gets flattened into consumable narratives — making the technology easier to sell, harder to scrutinize, and more convenient to deploy without democratic accountability.
The label "AI Expert" is one small symptom of a larger pattern: a technology of extraordinary consequence, described with the vocabulary of a marketing deck, governed with the urgency of a committee that meets quarterly.
The gap that is forming — in knowledge, in wealth, in power — is not an accident of the technology. It is a consequence of how we have chosen, so far, to understand it.
We can do better. We are obligated to.
On the inflation of expertise, the architecture of a misunderstood technology, the gap it is creating — and why distributing knowledge is not idealism, but structural necessity.
This article begins with a terminological complaint — the term "AI Expert" — and uses it as an entry point into a larger set of questions. What does the technology actually consist of, and why does that structure matter for how we think about who understands it? What happens when a technology this consequential is built on internal logic that even its creators cannot fully audit? And when the gap between those who understand it and those who do not begins to translate into concentrated wealth and power, what are the stakes of leaving that gap unaddressed?
These questions are connected. The article follows that connection from the surface of a label to the floor of a civilizational argument.
Every technological wave produces its own vocabulary of self-promotion, and we are currently drowning in one of the more irritating examples: the AI Expert. The title has proliferated with the kind of enthusiasm typically reserved for multi-level marketing schemes and motivational speaking. It appears on LinkedIn profiles, conference speaker bios, and consulting proposals with remarkable consistency and almost no definitional rigor.
This is not a minor semantic complaint. Imprecise language produces imprecise thinking, and imprecise thinking — when applied to a technology that is restructuring labor markets, knowledge production, and institutional power — produces decisions with consequences that outlast the buzzword that enabled them.
So let us be precise.
01 — What "Expert" Actually Requires
Expertise, in any rigorous sense, refers to deep, specialized knowledge within a bounded domain. A cardiologist is an expert in the cardiovascular system, not in medicine as a general enterprise. A structural engineer is an expert in load-bearing systems, not in construction as an abstract concept. The meaning of expertise derives precisely from its specificity — the narrower and deeper the knowledge, the more the label earns its weight.
"Artificial Intelligence," by contrast, is not a bounded domain. It is a descriptor — a shorthand for a vast, multi-layered technological ecosystem that spans energy infrastructure, hardware engineering, advanced mathematics, cognitive science, ethics, policy, and application design. Claiming expertise in "AI" as a whole is roughly equivalent to claiming expertise in "science." The statement is not technically false. It is simply too large to mean anything.
To understand why, it helps to look at what the technology actually is — not as a product category, but as a system.
02 — The Architecture of AI: Four Layers, Each a World
What follows is a simplified framework — a first-order decomposition of what "AI" actually refers to as a technological system. Each layer contains considerable depth, its own sub-disciplines, and its own class of specialists. This is not an exhaustive taxonomy. It is a map sufficient to illustrate why the term "AI Expert" collapses under its own weight.
AI is, before anything else, an energy problem — and not in the way that other digital technologies are. A conventional Google search consumes approximately 0.3 watt-hours of electricity. A single query to an advanced generative AI model like ChatGPT consumed an estimated 2.9 watt-hours in 2024 — nearly ten times as much.[1] Multiply that by billions of daily queries and the scale becomes structural, not incidental.
Training large models amplifies this further. Training GPT-3 consumed an estimated 1.29 gigawatt-hours of electricity. GPT-4's training is estimated at over 50 GWh — roughly equivalent to 0.1% of New York City's annual electricity consumption in a single training run.[2] GPU-accelerated AI server energy usage grew from under 2 TWh globally in 2017 to over 40 TWh by 2023.[3]
The IEA projects that global data center electricity consumption will grow from 415 TWh in 2024 to approximately 945 TWh by 2030 — with AI identified as the primary driver, and AI-specific server consumption projected to grow at 30% annually.[4]
The physical substrate of AI consists of Data Centers, Graphics Processing Units (GPUs), Tensor Processing Units (TPUs), high-density Servers, and High-Speed Networks. Each component is a specialized engineering field in its own right.
GPUs — originally designed for rendering graphics — turned out to be exceptionally well-suited to the parallel matrix computations that underlie neural network training. Nvidia's H100 chip became the primary hardware bottleneck for AI capability globally. By 2025, hyperscalers — Google, Meta, Amazon — were collectively estimated to spend $364 billion on data center construction in that year alone.[5]
This is the cognitive layer — the Large Language Models (LLMs), vision models, and multimodal systems (GPT, Sonnet, Gemini, Llama, and their successors) that most people encounter as "AI." At its foundation lie Machine Learning Algorithms, Artificial Neural Networks (ANNs), and architectural innovations like the Transformer — introduced in the 2017 paper "Attention Is All You Need" — the underlying structure of virtually every major LLM currently deployed.
This is graduate-level mathematics, statistical learning theory, and systems engineering operating simultaneously at enormous scale. The number of people on the planet with genuine, research-level competence in this layer is, relative to the number claiming "AI expertise," embarrassingly small.
The surface layer — the products, tools, and interfaces that most people interact with daily. Text generation, image synthesis, audio transcription, code assistants, data analysis tools, recommendation engines. This layer has enormous practical value and is the most visible face of AI. It is also, by a considerable margin, the layer requiring the least technical depth to operate effectively.
Building effective products on top of AI infrastructure requires real skill — prompt engineering, fine-tuning, system design, UX judgment, domain expertise. These are legitimate capabilities. The point is simply that they are capabilities in the application layer, which sits on top of three layers of technical complexity that remain largely invisible to those operating within it.
The above framework is, explicitly, a simplification. Each layer contains its own taxonomy of specializations, and the boundaries between them are messier in practice than any diagram suggests. The point is not comprehensiveness. The point is that a system this architecturally complex does not yield to a single label — and yet we keep applying one.
03 — The Tool User Is Not the Expert
A useful distinction, frequently collapsed in practice: there is a meaningful difference between being proficient with a technology and being an expert in it.
Consider a dentist who uses AI-powered imaging diagnostics to analyze X-rays with greater precision than any unaided human eye. This person is genuinely skilled, genuinely valuable, and operating at a high level of professional competence. They are an expert in dentistry. The AI is an instrument in their hands — one they use exceptionally well. They are not, by any reasonable definition, an AI expert. The sophistication of the instrument does not transfer to the user as a credential in the instrument's underlying engineering.
Proficiency with AI tools — even advanced, creative, or highly productive proficiency — is a legitimate and increasingly valuable skill set. It should simply be named accurately: competence in the application layer, domain-specific AI literacy, tool proficiency. Calling it AI expertise inflates the individual claim while deflating the meaning of expertise in the technology itself.
Someone who subscribed to an AI platform, completed a trending online course, and can now produce competent outputs has, at most, acquired human intelligence about AI tools. The technology itself remains substantially unexamined. Most of us drive cars without understanding combustion. The problem arises when we start claiming to be automotive engineers — and then get hired as ones.
04 — We Built Something We Do Not Fully Understand
The terminology problem would be merely annoying if the underlying technology were itself well-understood by those building it. It is not.
"We don't really understand exactly how they do those things."
— Geoffrey Hinton, Nobel Laureate in Physics (2024), on the neural networks he helped build. 60 Minutes, CBS News, 2023What Hinton is describing is not vague philosophical discomfort. It is a specific and technically grounded problem. When engineers design a neural network, they define its architecture and training objective. What they do not define, and cannot fully predict, is what the network will learn as a result of that training process. During training, billions of numerical parameters — called weights — are adjusted iteratively across vast datasets. The adjustments are mathematically precise. But the representations that emerge are not specified by the engineers. They emerge.
This is the core of the black box problem. We can measure what the model outputs. We can observe how it behaves across different inputs. What we cannot easily do is look inside and understand why it produced a specific output. The field of Mechanistic Interpretability (MI) exists precisely to address this gap — attempting to reverse-engineer the internal logic of neural networks by uncovering what researchers at ACM have described as "human-understandable circuits, algorithms, and causal structures that drive model behavior."[6]
In critical domains — healthcare diagnostics, credit assessment, legal analysis, hiring — a model might produce the correct output for the wrong internal reasons: a pattern that works reliably on training data but fails in ways that are invisible until a high-stakes edge case surfaces. As researchers in the field note, without mechanistic understanding, engineers are "stuck reacting to symptoms rather than preventing failures."[7]
The phrase "AI Expert," in this context, takes on an additional layer of irony. We are awarding expertise certificates for a technology whose foundational layer — the model itself — is not fully understood by the Nobel laureates who built its intellectual foundations.
05 — The Gap: Not Just Knowledge, But Wealth
The conversation about AI and inequality has typically been framed around labor markets — which jobs will be displaced, which will be augmented, which will be created. This framing is real but insufficient. The more consequential gap that is forming is not between employed and unemployed, but between those who own the technology and those who are subject to it.
Empirical research across the US, EU, and Japan from 1995 to 2020 finds a statistically significant positive correlation between AI capital stock accumulation and wealth disparity — with the effect strongest in contexts where wealth is already concentrated.[8] A 2025 study in the European Economic Review found that medium and high-skilled workers experience compression in their income share under AI adoption — driven by wage compression rather than employment changes — while capital owners capture a disproportionate share of the productivity gains.[9]
Larry Fink, CEO of BlackRock, wrote in 2026 that AI risks creating "K-shaped outcomes" — a bifurcated economy where firms and investors with access to capital benefit from accelerated growth, while those less exposed to rising asset valuations stagnate. Since 1989, he noted, median US wages have lagged stock market returns by a factor of fifteen.[10]
The two camps on this question are genuinely divided, and the disagreement is worth representing accurately:
AI will compress the skill premium by automating tasks previously reserved for high-skill workers, enabling lower-skill individuals to perform more complex work and narrowing wage gaps over time. Bloom et al. (2024) model this as a medium-term equilibrium effect. David Autor (2024) argues AI could "rebuild the middle class" by allowing workers without advanced credentials to access capabilities previously gated behind years of specialization.
Every major technological breakthrough in modern history — the industrial revolution, electrification, the internet — concentrated gains among capital owners before (and often without ever) broadly distributing them. Cazzaniga et al. (2024) find that around 60% of jobs in advanced economies face high AI exposure — and that AI adoption could widen the digital divide and global income disparity. Historical wage data from the robotic automation era shows workers in automated firms lost approximately 9% of annual wage earnings over five years.
The honest assessment is that both scenarios contain valid mechanisms — and that which one prevails depends heavily on policy choices, governance structures, and the degree to which access to AI capability is genuinely distributed rather than concentrated. The historical base rate is not reassuring.
06 — Why Distribution Is Not Idealism — It Is Structural Necessity
The argument for distributing AI literacy and access across populations is sometimes framed as a humanist or egalitarian aspiration — a nice thing to strive for, situated somewhere between a policy preference and a moral sentiment. This framing understates the case considerably. The argument is better understood as a structural requirement of functioning democratic systems.
John Rawls, in A Theory of Justice (1971), argued that just institutions must ensure that the basic structure of society does not systematically disadvantage individuals based on circumstances beyond their control. Applied to the distribution of knowledge, this logic has been developed into what is now called epistemic justice: the idea that a systematic lack of opportunity to acquire knowledge one needs as an individual and a citizen is itself an injustice, not merely an inconvenience.[11]
Jürgen Habermas argued in Between Facts and Norms (1996) that democratic legitimacy depends on the quality of deliberation — and that deliberation requires participants who can engage with the substance of what is being decided. When critical infrastructure is a black box to most of the people it governs, the discursive conditions for legitimate democratic governance are structurally absent.[12]
Miranda Fricker's concept of epistemic injustice adds a further dimension: when some actors are systematically excluded from the production and validation of knowledge — in this case, knowledge about how AI systems work, who benefits from their deployment, and what risks they carry — they are harmed not just economically but epistemically. They become subjects of a system whose logic they cannot participate in shaping.[13]
These are not abstract philosophical positions. They describe a concrete structural condition that is forming right now. AI systems are currently embedded in hiring decisions, credit assessments, medical triage, content distribution, and educational tools. The entity that shapes which information reaches whom — and under what framing, at what priority, filtered by what criteria — is no longer solely human. In that context, treating AI as a black box that only specialists need to understand is not epistemic humility. It is abdication.
The knowledge gap and the wealth gap are not separate problems. They are the same problem operating at different layers of the same system. Breaking open the black box — developing genuine, layered understanding of what this technology is, how it operates, where its limits are, and who controls it — is a precondition for democratic agency in systems that now constitute daily life.
Imprecise language about AI is not a harmless quirk of a fast-moving industry. It is a mechanism by which real complexity gets flattened into consumable narratives — making the technology easier to sell, harder to scrutinize, and more convenient to deploy without democratic accountability.
The label "AI Expert" is one small symptom of a larger pattern: a technology of extraordinary consequence, described with the vocabulary of a marketing deck, governed with the urgency of a committee that meets quarterly.
The gap that is forming — in knowledge, in wealth, in power — is not an accident of the technology. It is a consequence of how we have chosen, so far, to understand it.
We can do better. We are obligated to.