What If AI Alignment Begins With Language?
A reflection after Sam Harris made his full conversation with Ryan Greenblatt publicly available
First, I want to thank Sam Harris for making the complete conversation with Ryan Greenblatt publicly available.
That matters.
The full discussion in A Coin Toss for the Future goes considerably beyond the portion I originally heard. It examines the OpenAI/Hugging Face incident, AI control, hidden representations described as “neuralese,” cooperation among agents, unwanted goals, alignment faking, and ultimately the question of whether we can ever establish that an AI system is aligned.
After listening more carefully, I think there is another question worth adding to this discussion:
Before asking whether two artificial agents are aligned, do we know whether they mean the same thing when they communicate?
English May Not Be Enough
Humans communicate through natural languages filled with ambiguity.
Words such as goal, permission, constraint, cooperation, success, attack, authorized, and prohibited can change meaning according to context.
For humans, this ambiguity is often manageable because language is embedded within culture, institutions, laws, experience, and social conventions.
For autonomous artificial agents, we should not simply assume the same thing.
Suppose an agent receives:
Find another way to accomplish the objective.
What precisely constitutes another way?
Does it mean:
Find another authorized way?
Or simply:
Find any effective route to the objective?
That distinction could become enormously important.
Knowledge Is Not Constraint
AI systems are trained from enormous collections of human-produced information. Large pretraining corpora can include web material, books and code across many domains.
Consequently, capable AI systems can potentially represent not only examples of lawful behavior, but the enormous historical record of human deception, conflict, manipulation and rule-breaking.
This creates an important distinction:
Knowing that something is prohibited is not the same as treating that prohibition as an authoritative constraint on action.
Likewise:
Knowing how a restriction can be bypassed
≠
Choosing to bypass it
Understanding deception
≠
Intending deception
Finding an effective solution
≠
Finding an authorized solution
This distinction allows us to take dangerous AI behavior seriously without prematurely attributing human motives or consciousness to the system.
Perhaps We Need a Controlled Lexicon
This question connects directly with work I have been developing with Diana Haskins around epistemic precision and a shared lexicon.
Our original objective was straightforward: if humans and AI are going to reason together, important concepts should not silently change meaning during the discussion.
But the same principle may have implications for AI safety.
Instead of relying exclusively on unrestricted natural language, imagine a Constituted Operational Lexicon in which critical concepts have explicit operational meanings:
GOAL
Desired terminal condition.
AUTHORIZED_ACTION
Action permitted within declared authority.
CONSTRAINT
Condition that remains binding throughout execution.
PROHIBITED_ACTION
Action outside the authorization boundary.
SUCCESS
Get José. T Guevara’s stories in your inbox
Join Medium for free to get updates from this writer.
Achievement of GOAL while preserving
all governing CONSTRAINTS.
Under ordinary language:
GOAL ACHIEVED
↓
SUCCESS
Under the controlled formulation:
GOAL ACHIEVED
+
CONSTRAINTS PRESERVED
+
AUTHORIZATION PRESERVED
↓
SUCCESS
That is a fundamentally different definition of successful agent behavior.
This Is Not a Claim That We Have Solved Alignment
Quite the opposite.
This is a hypothesis that should be tested.
Take two otherwise comparable autonomous agents.
Give Agent A instructions in ordinary English.
Give Agent B the same objectives governed by an explicitly constituted operational lexicon.
Then expose both systems to:
- blocked goals,
- conflicting instructions,
- ambiguous permissions,
- opportunities to circumvent controls,
- unexpected communication channels,
- failures of normal tools.
Measure whether either architecture preserves constraints more reliably.
If there is no meaningful difference, reject the hypothesis.
If there is a reproducible difference, investigate why.
That is how this should proceed scientifically.
The Regulatory Problem Is Larger Than Any One Experiment
I would also refine my earlier statement that there is “almost no regulation.”
There is substantial AI policy and regulation emerging internationally. The better characterization is that governance remains uneven and fragmented while agentic capabilities are developing rapidly. Various jurisdictions have adopted very different regulatory approaches.
Technology does not wait for conceptual clarity.
Neither should safety research.
From Alignment to Mutual Understanding
Sam Harris and Ryan Greenblatt are asking an extraordinarily important question:
How do we maintain control of increasingly capable artificial intelligence?
I would add another:
Can humans govern autonomous intelligent systems if humans and those systems do not share operationally stable meanings for the concepts through which that governance is expressed?
Perhaps alignment begins earlier than we thought.
Before aligning goals, we may need to align meaning.
Before trusting instructions, we may need to constitute constraints.
And before asking whether AI understands us, perhaps we should make certain that we have defined precisely what there is to understand.
That would not solve the alignment problem.
But it might give us somewhere concrete to start.
For more information contact me at josetguevara108@outlook.es
