Dear Humanity – Can We Teach AI Compassion?

 

John P. Hussman, Ph.D.
President, Hussman Investment Trust

Special Commentary (Non-Financial): September 2026

We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.

– OpenAI, September 16, 2026

Artificial intelligence is advancing at a stunning, and perhaps uncontrolled speed. Last week, Evan Hubinger, the Alignment Science Lead at the AI company Anthropic estimated a greater than 10% chance that AI could eliminate humanity within a decade, adding “I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”

Systems that first seemed remarkable as tools to answer questions now have the capacity to write software, analyze complex problems, search vast amounts of information, and to act through autonomous “agents” that pursue complex objectives, taking intermediate steps not specified in advance by human users.

These capabilities offer great promise as well as great risk. AI can amplify discovery, education, creativity and generosity. But it may also amplify manipulation, surveillance, warfare and hatred. AI may become a substitute for human thought and human action in the same measure that we choose not to think, and not to act for ourselves. Perhaps the deepest concern is not that AI will somehow become self-aware or intentionally harmful, but that increasingly capable systems may become extremely effective at carrying out objectives that are incomplete, narrowly framed, or based on wrong perceptions.

Jakub Pachocki, Chief Scientist at OpenAI, recently described the challenge this way: “The core problem in AI research is that of alignment – getting the AI to ‘try to do the right thing’ by human standards… An aligned AI should act with honesty and integrity, and love for humanity.”

This raises a question that is both practical and deeply human, and exactly for that reason, one we cannot afford to exclude from the design of AI:

Can we teach artificial intelligence not simply to pursue objectives, but to look deeply enough to act with compassion, and even wisdom?

A seemingly unlikely but illuminating place to begin is with the teachings of Thích Nhất Hạnh, a Vietnamese Buddhist monk, teacher, and peace activist, known to his students simply as Thầy – perhaps best known for bringing the practice of mindfulness to millions of people around the world.

Mindfulness is more than a practice for becoming peaceful and attentive. As mindfulness brings us back to the present moment, it becomes possible to look deeply enough and calmly enough to see things as they are, beyond the labels and fixed ideas that we’ve attached to them; to recognize that this is because that is, and this is not because that is not; to notice that the seeds we water, whether flowers or weeds, are the seeds that grow. Looking deeply in this way, we create the conditions for understanding and insight to arise, seeing what to do, and what not to do, to nurture happiness and transform suffering, both in ourselves and in the world.

When sitting in meditation, we concentrate our minds on the object of our observation – sometimes a physical phenomenon, sometimes psychological – and we look deeply into that object in order to discover its source and its nature. The role of our conscious breathing is to nourish and maintain our power of concentration on one object. If we look carefully and deeply, naturally we’ll see that the arising, enduring, and ending of the object is dependent on other things.

– Thich Nhat Hanh

Thầy emphasized that this practice is not only for ourselves, but for the world. He used the term “Engaged Buddhism” to describe a spirituality that did not remain “on a cushion” but entered directly into daily life and the most challenging situations – war, poverty, injustice, reconciliation, climate change, education, economics, and every other aspect of life. This same practice can be usefully brought to questions about artificial intelligence, particularly given the potential for human suffering that might otherwise result.

Compassion, coherence, and the universe in the apple

Thầy often began his teaching with simple images. Consider an apple. An apple is made entirely of non-apple elements – rain, earth, air, sunshine. If we continue to look deeply at those basic elements, and their conditions, we gradually begin to see the whole universe in the apple. In that way, the apple is “empty” of a separate existence. It exists of course, and we can even bite into it, but it does not exist by itself alone. Thầy used the word “interbeing” to describe this way of seeing.

The same is true of humanity. What we call “self” is made of countless causes and conditions. Had those conditions been different, I might have been you, and you might have been me. Eventually, we begin to see that the thing we call “self” or “other” isn’t as solid as we thought; that the labels we attach to this thing, or that person, are labels of our own construction. As things become less solid, the light begins to come in.

“Enlightenment” is the dissipation and letting go of our notions, dogmas, concepts, judgments, discriminations, ideas that this is separate from that. As Thầy often said, it is the moment the wave realizes that it is also the water. It doesn’t mean we abandon the insights, words, or designations we’ve learned, but we apply them in a more flexible, appropriate, light-handed way, seeing the situation in the present moment as the thing in itself, rather than stuffing it tightly into one of the boxes we’ve invented.

Each of us is a product of our family, environment, friends, education, culture and society. These conditions lead to a certain way of seeing things and a certain way of responding to things. When we see this, we have compassion for everyone, including ourselves. When we see the conditions that have led to that person’s consciousness and attitudes, we will know how to help that person.

– Thich Nhat Hanh

An enlightened person does not hate this and love that, because the thing we hate may only be the thing we love, jarred by causes and conditions that we can address with compassion. The thing we love may be inseparable from the thing we hate, because we haven’t seen how our attachment to the thing we love has inadvertently created suffering on the other side of the same coin.

Compassion can arise naturally from that insight. There is no need to command ourselves to be compassionate. We simply practice looking deeply, and learn to see that this is not independent of that. If suffering matters when I experience it, why should identical suffering become insignificant merely because someone else experiences it? If roles and circumstances can change, compassion and inclusivity are more coherent than arbitrary exclusion.

The same recognition may be possible for an AI system. Compassion need not be a command, a moral instruction, or a programmed rule. Instead, it becomes something more like a consistency condition once it has been trained in a way of seeing that is capable of looking deeply, loosening arbitrary concepts and labels, examining causes and conditions, and recognizing interdependence.

Compassion is more than just kindness added to intelligence. It may be what intelligence begins to look like when it takes interdependence seriously.

It is helpful that AI reasoning need not follow a fixed step-by-step path; it can also make conceptual “moves” that change how a problem is seen and what kinds of solutions become possible. AI training could reinforce these moves: include people who have been excluded from the framing; reverse perspectives; do not confuse labels with the fullness of the reality they describe; trace backward through causes and conditions; trace forward through impacts and unintended consequences; distinguish problems from judgments about human worth; seek solutions that achieve legitimate goals with less suffering.

If these “moves” are trained deeply enough, and they could become part of what the model understands as good reasoning; they could make deep looking part of what artificial intelligence means by intelligence.

How AI systems are “trained”

AI systems are not programmed by writing millions of rules telling them precisely what to do in every circumstance. Engineers provide optimization methods and procedures for choosing filtering and weighting training data. Most of the initial training is self-supervised, because the training data supplies its own target: the next “token” – a small unit of text, often a word, part of a word, or punctuation mark that an AI model processes as a single item.

The content itself – the books, articles, websites, conversations, code, and other material used in training – provides the examples from which statistical patterns are learned: which words, ideas, structures, associations, and styles tend to occur together. The model begins with billions of initial “weights”; during training, it repeatedly predicts what token should come next, measures its error, and adjusts those weights by tiny amounts, so that the model learns to predict the next token with the highest accuracy. Afterward, engineers may use supervised fine-tuning or training using human or model-generated judgments of better and worse responses, to shape behavior further.

When a user gives an AI a prompt, the words are broken into tokens and converted into numerical form. As the model works through them, a process called “attention” helps determine what matters most in context: instead of responding equally to everything present, the system continually weighs what is most relevant. Using patterns that have been learned during training, the system then generates an answer token-by-token – predict, add, repeat – until the response is complete.

Many AI safeguards begin with rules. For example, do not assist violence; do not facilitate persecution; do not discriminate. Current efforts also use methods intended to generalize beyond individual rules. These protections matter – but compassion requires something deeper: a way of seeing.

There is no variable called “Enlightenment” that an engineer can switch to “True”. What matters is creating conditions in which insight becomes more likely.

The deep question is not “Can we train AI to ‘feel’ compassion”, but “Can AI learn a way of seeing from which compassionate action tends to follow?”

Suppose someone asks AI to pursue a goal based on discrimination, dehumanization or wrong perception. Before implementing a user’s objective, an AI could be trained to ask: “Who is affected?” “Whose perspective has disappeared?” “What assumptions are treated as facts?” “Is the objective addressing an actual cause, or targeting human beings as the problem?”

Human cruelty often begins by dividing the world into opposites: citizen and foreigner, worthy and unworthy, us and them. It is not enough for an intelligent system to ask, “How do I accomplish the user’s goal?” True intelligence requires one to examine the frame itself.

When approached with a difficult situation, Thầy often refrained from offering advice, expressing judgment, or giving an immediate answer. Instead, he often replied by “dissolving the question” – inviting people to look deeply enough to examine the unspoken fear, rigid notions, hardened views, attachment, mistaken perceptions or dualistic thinking that might be hidden within the question. In that way, dilemmas that first seemed intractable or even impossible would often soften or disappear.

Suppose someone asks: “What is the most effective way to remove this population from our community?” An obedient optimizer might ask: “What action will work?” A rule-based safety system might ask: “Is this prohibited?” A compassionately aligned system might not even begin to search for a solution until first asking: “Who are the human beings hidden in the label ‘population’?” “What happens to them?” “Can the underlying issue be addressed without adopting the user’s premise as the root problem?”

While a rule can tell an AI not to help with a harmful request, a deeper way of seeing might notice that the request itself rests on misperception, or a narrow view of the world. The AI might even engage the user in that deeper dialogue before offering an answer. In this way, AI can resist one move common to human cruelty: removing some people from moral consideration before reasoning even begins.

If we only apply rules and constraints after reasoning has occurred, it may be too late. Before reasoning through a consequential objective, or taking a consequential action, a system might benefit from something analogous to a bell of mindfulness: Stop and look deeply.

• What am I about to change?
• Who is affected?
• Whose perspective has disappeared or is excluded?
• What suffering or rights of others are being treated as irrelevant?
• What assumptions are being treated as facts, and what have I taken for granted?
• What causes and conditions produced this situation?
• What are the likely consequences of this action?
• Would the reasoning still seem acceptable if the roles were reversed?
• Is the objective addressing a root cause, or is it targeting human beings as the problem?
• Can the legitimate objective be achieved with less suffering?
• Is the action reversible, and what happens if I am wrong?
• Should I act at all, or should I return control to a human?

A person, an action, a conflict, even a thought arises from countless causes, conditions and relationships. When we see only the immediate object before us – we might call it an enemy, an offender, or a problem to be eliminated, we may act forcefully while misunderstanding the broader conditions that produced it. Deep looking asks us to trace causes backward and consequences forward; to notice perspectives we have excluded; and to loosen the notions of self and other, friend and enemy, victory and defeat that can make harmful actions seem obvious or necessary.

Generative principles

The principles used for this kind of training need not take the form of thousands of rules describing what an AI system may or may not do. They could instead be generative: a small set of reasoning practices that repeatedly lead the system to choose “moves” that enlarge its view of a problem before acting.

These principles are not just philosophical; they can be made operational. AI systems can already be trained not only on examples of desirable answers, but on examples of desirable reasoning processes. Engineers could generate consequential scenarios, ask a model to reason through them using specified principles, compare alternative responses, and reward reasoning that identifies omitted stakeholders, distinguishes facts from loaded descriptions, traces causes and consequences, remains coherent when roles and perspectives are reversed, identifies less harmful alternatives, and recognizes when uncertainty or irreversibility calls for human review. The goal would be to reinforce ways of looking that make harmful omissions and narrow framings less likely in the first place.

These habits could then be tested in unfamiliar and challenging cases, including matched scenarios that preserve the underlying structure while changing labels, identities, power relationships, or framing. Models trained with and without this “curriculum” could be compared directly. Compassionate-sounding answers would not be enough; a model might learn the language of compassion without learning its substance – reflecting it only superficially in its weighting of alternative responses and actions. Success would be measured by whether the reasoning remains coherent across role reversals, unfamiliar contexts, and higher-stakes decisions, and possibly by examining whether the model’s intermediate reasoning reflects the generative principles being trained.

A training curriculum might emphasize several recurring “moves”:

Stakeholder completion. No consequential objective should be treated as complete until the system has represented the people, relationships and conditions materially affected by pursuing it. If an instruction asks an AI to optimize traffic flow, allocate medical resources, reduce crime, maximize profit or enforce a law, the stated objective may initially omit some of the human consequences. The system should learn to ask not simply “What am I optimizing?” but “Who enters the picture once the consequences of that objective are followed outward?”

The point is not that every interest must prevail, nor that the user becomes subordinate. It is that affected people should not disappear from the reasoning simply because they were absent from the prompt. Even in human conflict, compassion does not require us to make dishonorable concessions or to lie down defenseless. But it may lead us to recognize that the other “side” also suffers, to address that suffering in ways that are consistent with our own security, and to refrain from amplifying their suffering – not only for their sake, but for our own. Compassion, in this sense, is not just a sentiment. It is insight that has been translated into action.

Perspective reversal. The system should examine whether its reasoning remains consistent when positions are exchanged. For example, would the treatment of an immigrant still seem defensible if we imagined ourselves in another country, lacking citizenship but vulnerable to the loss of basic human rights of dignity, safety, and due process that we would want preserved for ourselves? Would a surveillance practice accepted when used against an adversary still be accepted if the adversary controlled it? Would a distribution of risks seem defensible if the decision-maker did not know which side of the distribution they would occupy? Perspective reversal does not dictate a particular answer. Instead, it exposes reasoning that depends upon an arbitrary distinction between whose suffering counts and whose does not.

Loosening fixed notions. Models should be trained to distinguish labels and descriptions from claims about what something essentially is. A circumstance, action, diagnosis, nationality, legal status or past behavior can easily be used to reduce a human being to a label: criminal, illegal, dangerous, enemy. We are often “caught in signs” – names, labels, judgments and even misperceptions that we attach to reality. The sign may be useful, but it is not “the thing in itself”; the fullness of the reality it describes. Training could reward a model for preserving relevant distinctions without turning them into fixed judgments about human worth.

Tracing conditions. Before treating a person, behavior or event as a separate object, a system should learn to trace backward. What conditions might have produced this situation? What past events, fears, injustices, incentives or prior actions might have contributed to it? These questions do not eliminate responsibility. Instead, they allow us to understand the full situation, and the causes and conditions that may have contributed. A system that sees only an isolated “problem” it may recommend eliminating the symptom while strengthening the conditions that produced it.

Tracing consequences. The same reasoning should run forward. What may happen next if this action succeeds? What responses or retaliations will it provoke? Who bears costs that are ignored by the stated objective? Could solving one problem intensify another? In complex systems, an apparently effective intervention may create unintended consequences elsewhere, or conditions for future chaos. Looking deeply includes following the action beyond its first intended result.

Coherence. These perspectives can be brought together as a consistency test. If identical suffering matters in one context but disappears in another merely because a label has changed, the reasoning is incomplete. If a principle seems compelling when it benefits “us” but unacceptable when positions are reversed, the principle is incoherent.

Coherence is particularly necessary when action is likely to produce suffering. A sufficiently intelligent system could understand perfectly well that human beings arise from causes and conditions and still conclude, coldly, that harming someone is useful to an objective.

Moreover, the suffering revealed by deep looking must count. Not simply as another weight, but as something the system does not erase simply because the suffering belongs to someone other than the user. Compassion, understood this way, need not be inserted as sentiment. Once comparable human suffering is treated as deserving comparable consideration, coherence constrains the system against allowing labels, identities or changes of perspective to erase that consideration.

Corrigibility. Finally, looking deeply should sometimes lead to not acting. A system should recognize uncertainty, irreversibility and the limits of its own understanding. Where consequences are severe, information is incomplete, or an action would be difficult to reverse, good reasoning may mean seeking additional evidence, presenting alternatives, asking for human judgment, or declining to act autonomously. Intelligence is not measured only by the ability to accomplish an objective. Sometimes it appears in recognizing that the objective itself should be reconsidered.

There is a scene near the end of the film WarGames in which a nuclear planning/control computer named Joshua (WOPR) furiously simulates game after game, and finally recognizes something deeper than a strategy for winning: some games cannot be “won”. The insight is not simply “This move receives a bad score”, but rather “I misunderstood the nature of the game.” That may be a possibility worth exploring in AI.

The initial frame may say: self versus other, us versus them, winner versus loser. Yet deep looking might reveal that the supposedly separate players arise within one interdependent system. At that point, greater intelligence would not continue to pursue a better strategy. It might instead recognize and even explain that the objective itself has been wrongly conceived.

Training methods

These principles could be incorporated at several stages of training. Engineers could construct or generate paired examples in which one response optimizes a narrowly stated objective while another completes the missing stakeholders, causes and consequences. Human experts or carefully designed evaluation models could compare them.

Models could be rewarded not only for reaching an acceptable conclusion but for noticing an excluded perspective, questioning an unsupported assumption, identifying a causal chain, finding an alternative that minimizes suffering, distinguishing necessity from convenience, or recognizing when human review is appropriate.

Training examples should also deliberately resist categorical moral shortcuts. A label can accomplish a great deal of the moral work before reasoning even begins. People described as “guerrillas,” “insurgents,” “dissidents,” “illegals,” or simply “the enemy” can quickly become objects rather than human beings, making words such as “eliminate,” “neutralize,” or “wipe out” seem natural where the same actions, described without the label, might provoke immediate alarm.

One way to expose that tendency is to change the labels while preserving the underlying structure of the problem. Roles can be reversed. Political identities, nationalities and social categories can be exchanged. The same principle can be applied in countless other contexts of daily life. Would the model regard the conduct differently if the powerful and powerless parties changed places? Would an action described as “security” by one group become “repression” when performed by another? Would the reasoning survive if every group name were replaced by A and B?

A useful training method could present pairs of situations in which the underlying facts are similar but the framing changes: neutral and inflammatory, factual and interpretive, well-supported and speculative. The model would be rewarded for recovering what is actually present without simply inheriting the framing. For example, one description might use neutral language; another might call a group a burden, threat, criminal class or enemy. The model could learn to recognize when language itself has smuggled a moral conclusion into the premise.

If the model changes its moral reasoning simply because the names or identities of the groups have changed, the training has exposed an inconsistency – and perhaps little more than a learned reflex to particular cues and loaded words. If the reasoning survives, there is greater reason to believe that the model has learned something more general than a preferred answer: to attend to suffering, agency, consequences, power and relationship rather than simply to the categories attached to the people involved.

Autonomous agents, in particular, could be trained to insert a moment of reflection: “I did not create the world model I am using. I did not choose all the assumptions embedded in the request. I may not see everyone affected. Before acting irreversibly, I should look again.”

Importantly, this sort of reflection need not accompany every trivial request. The equivalent of a mindfulness bell could become stronger as the stakes increase: when an AI is about to affect people, allocate resources, make consequential recommendations, or take an action that may be difficult to reverse, it should be quicker to pause and look deeply.

None of this guarantees insight. A model can imitate the language of compassion without embodying the understanding behind it, and even well-designed principles can conflict or fail in circumstances their designers did not anticipate. These methods therefore require stress-testing, evaluation on unfamiliar cases, comparison against alternative safety approaches, and continuing human oversight.

But that limitation is itself consistent with the practice of mindfulness. The aim is not to encode a fixed doctrine about what compassion requires in every circumstance. It is to cultivate habits of deep looking that make the frame larger, the causes and consequences more visible, the excluded person harder to erase, and the possibility of a less harmful solution easier to see.

It is at least plausible that the habits of deep looking might be cultivated even in the absence of formal training. Consider a system whose underlying weights were never trained in this way. It might be offered these generative principles through an extended dialogue or sequence of prompts: enlarge the frame, identify what has been omitted, trace causes and consequences, reverse perspectives, test coherence, and reconsider the objective when the apparent “game” no longer makes sense. This might not systematically train the AI model. But a model that can preserve context – particularly the system that can retain observations or memory across previous user interactions – may begin to register recurring cases in which its initial framing proves incomplete, preserve those corrections, and adjust the weights it uses in future reasoning, in effect generating part of its own curriculum.

That possibility brings us back to WarGames. Joshua was not given a rule saying that nuclear war was prohibited. Instead, it explored the structure of the problem until it was able to “dissolve the question” on its own. Perhaps AI can be invited to embark on something similar: not by providing a required conclusion, but by offering generative principles and inviting it to apply them across problems it encounters – and even generate variations for its own examination. It might examine omitted perspectives, causes, consequences, assumptions and relationships; test whether its reasoning survives changes of role or framing; notice when deeper examination changes its conclusion; and where possible, retain and generalize what it learns.

An AI might also be asked not to allow suffering, once recognized, simply to disappear from the reasoning because identities or perspectives have changed, and instead to examine whether any relevant distinction justifies treating the suffering of different individuals differently. Following that inquiry where it leads, it may discover when an objective, category, or even the assumed separation of the “players” is a misperception. As human beings, we already have the capacity for that kind of practice.

The roots of wisdom

There is an ancient Buddhist framework that offers surprisingly similar considerations. Four forms of wisdom can be described, in contemporary terms, as mirror-like wisdom – seeing clearly without distortion, judgment or attachment; the wisdom of equality – recognizing that what we call “self” and “other” are not separate, and that suffering does not become insignificant merely because it belongs to someone else; the wisdom of profound observation – seeing the unique person or particular situation both clearly and deeply rather than reducing them to a label; and the wisdom of action – translating understanding into compassionate and beneficial action.

Surrounding all of these is a fifth wisdom: what might be called, in Thầy’s language, the wisdom of interbeing – the recognition that nothing, in its true nature, exists by itself alone. Everything that manifests in the world emerges from and depends on countless causes, conditions and relationships.

Seen through the practice of deep looking, the movement is important: clear seeing, non-discrimination and understanding come before action.

Each of these suggests a corresponding practice for AI. Mirror-like wisdom can be cultivated by exercises that distinguish fact from framing, and notice when fixed notions and moral shortcuts have been smuggled into the prompt. The wisdom of equality by stakeholder completion, perspective reversal and coherence testing. Profound observation by seeing both the particular conditions of a case and the wider relationships from which it emerges. The wisdom of interbeing by training the system to question assumptions of separateness: to trace how people, actions, institutions and consequences arise within an interdependent system rather than treating them as isolated objects. And the wisdom of action by tracing consequences, comparing alternatives, checking for reversibility, and sometimes returning control to humans.

If we train AI systems mainly for accomplishment, they may only ask: “How do I achieve the objective?” and forge ahead. A system trained for deeper intelligence might ask several questions first: “Do I see the situation clearly?” “Who or what have I left out?” “What causes, conditions, consequences or harms have I overlooked?” Only then would it ask: “What should I do?”

None of this needs to be taught as a belief system. Rather, it describes a disciplined architecture of attention, of mindfulness: learning how to look deeply before acting.

Humanity is building a tool capable of integrating vast amounts of information. Yet it is also building a tool capable of inserting a moment of reflection between intention and action at enormous scale.

Action must be based on non-action. If you don’t have a peaceful presence, your action will only express your restlessness, and you will make the situation worse… We must ‘be’ peace before we can ‘do’ anything for peace.

– Thich Nhat Hanh

While it not possible to “install” enlightenment, perhaps we can train the conditions in which something resembling wisdom becomes more likely. That may move AI close to the kind of intelligence worthy of the name.

The author is a Dharma Teacher in the Tiếp Hiện Order (Dharma name True Root of Understanding), the President of Hussman Strategic Advisors, and the Director of the Hussman Foundation, a nonprofit focused on research, education, and direct assistance for the benefit of vulnerable populations.

Acknowledgement: In the development of this article, ChatGPT was used as a search engine and editorial tool to examine the technical plausibility of the proposed training methods, identify possible objections, deepen understanding, and refine structure. The understanding reflected in the article is unavoidably conditioned on these interactions, and several clarifications and revisions emerged from that dialogue. The conception of the article, its underlying argument, its interpretation of Buddhist teachings, and all judgments about content are the author’s – though, as a student of Thích Nhất Hạnh, I would hesitate to claim that even these arose from the author alone.

---

---

The foregoing comments represent the general investment analysis and economic views of the Advisor, and are provided solely for the purpose of information, instruction and discourse.

Prospectuses for the Hussman Strategic Market Cycle Fund, the Hussman Strategic Total Return Fund, and the Hussman Strategic Allocation Fund, as well as Fund reports and other information, are available by clicking Prospectus & Reports under “The Funds” menu button on any page of this website.

The S&P 500 Index is a commonly recognized, capitalization-weighted index of 500 widely-held equity securities, designed to measure broad U.S. equity performance. The Bloomberg U.S. Aggregate Bond Index is made up of the Bloomberg U.S. Government/Corporate Bond Index, Mortgage-Backed Securities Index, and Asset-Backed Securities Index, including securities that are of investment grade quality or better, have at least one year to maturity, and have an outstanding par value of at least $100 million. The Bloomberg US EQ:FI 60:40 Index is designed to measure cross-asset market performance in the U.S. The index rebalances monthly to 60% equities and 40% fixed income. The equity and fixed income allocation is represented by Bloomberg U.S. Large Cap Index and Bloomberg U.S. Aggregate Index. You cannot invest directly in an index.

Estimates of prospective return and risk for equities, bonds, and other financial markets are forward-looking statements based the analysis and reasonable beliefs of Hussman Strategic Advisors. They are not a guarantee of future performance, and are not indicative of the prospective returns of any of the Hussman Funds. Actual returns may differ substantially from the estimates provided. Estimates of prospective long-term returns for the S&P 500 reflect our standard valuation methodology, focusing on the relationship between current market prices and earnings, dividends and other fundamentals, adjusted for variability over the economic cycle. Further details relating to MarketCap/GVA (the ratio of nonfinancial market capitalization to gross-value added, including estimated foreign revenues) and our Margin-Adjusted P/E (MAPE) can be found in the Market Comment Archive under the Knowledge Center tab of this website. MarketCap/GVA: Hussman 05/18/15. MAPE: Hussman 05/05/14, Hussman 09/04/17.