betaFeedback

WorksComputer Science

The Tiglath Pileser principle and models of persuasion

Abstract

This paper develops a formal framework for analysing how agents extract reliable information from narratives and communicative interactions involving strategic deception. Building on Rubinstein and Glazer's models of persuasion, we introduce Tiglath-Pileser relations, a generalization of codex-based guidance relations, to formalize Baruch Halpern's Tiglath-Pileser principle: despite systematic bias, reported claims must retain a connection to the truth. We analyse these relations set-theoretically and axiomatically, identifying conditions for their successful application in persuasion. Our dynamic approach examines how such relations are formed, revised, and strategically manipulated, extending existing models beyond one-shot interactions. Two case studies, including a formal reconstruction of Halpern's interpretation of the biblical account of King David, illustrate the framework's application to historiographical reasoning. Combining logic, epistemic reasoning, and game theory, the paper provides an account of belief formation and truth approximation under partial information and strategic manipulation.

Introduction

A core question in historiography (as in many other academic and non academic fields) is: how much of a text (or of a story as is told in a text or other medium) can we believe? In his analysis of the story of King David as it is told in the Book of Samuel and 2 Kings [11], the biblical historiographer Baruch Halpern suggests (cf. [11 Part III, Ch. 5]) that this text1 can be analysed using the same basic interpretive strategy which applies to the Assyrian annals of Tiglath-Pileser, in which “the torching of a grain field is the conquest of a whole territory beyond it. A looting raid becomes a claim of perpetual sovereignty. But this does not mean that campaigns can be confected. The technique is that of putting extreme spin on real events.” Halpern continues (ibid.):

Interpreting such literature demands only a simple rule, the Tiglath-Pileser principle. The question is, what is the minimum the king might have done to lay claim to the achievements he

publishes? Looting a town? He shoplifted a toothbrush from the local drug store. Ravaging the countryside? Perhaps he trampled crops near a farmstead. Receiving submission from distant kings in lands one hasn’t invaded? A delegation arrived to inaugurate diplomatic relations. Each small mark of prestige becomes the evidence for a grand triumph. [...] The point is, such figures cannot make claims without any basis in fact [...] But it is always a temptation to paint one’s achievements in the best of all possible lights. To correct for this tendency, one has merely to imagine what insignificant action would produce the claim that “I conquered Egypt” (a raid on a border station in the Sinai). The same political hermeneutic applies reasonably well today.

Thus, the Tiglath-Pileser principle is grounded in the assumption that the given narration is anchored in truth (“this does not mean that campaigns can be confected”, “such figures cannot make claims without any basis in fact”), because the narrator must take into account what knowledge his original audience has that cannot be contradicted. Moreover, the Tiglath-Pileser principle provides a way of attaining a more accurate approximation of the truth, by taking into account both the original audience’s knowledge of the actual circumstances (which constrain the narrator to not make claims that would be immediately disbelieved), and the narrator’s agenda, and thus correcting for the tendency of this agenda to ‘spin the truth’. Building upon this ‘simple rule’ enables Halpern to reach the following surprising conclusions in respect to King David’s story [11 Afterword, p. 479]:

[...] the only murder in the books of Samuel of which [David] was probably innocent is the one murder of which he stands accused in the apology. [...] At the end of his life, David was certainly betrayed by one of his harem women, Bathsheba, and his foreign mercenary bodyguard. Instead of his heir, Adonijah, they installed Solomon, hitherto known as the son not of David but of Uriah the Hittite, on the throne. [...] this second king of Judah quickly cleaned house, erasing his rivals, and probably all of David’s sons, from the scene.

This paper presents a framework which, inter alia, can formally represent and trace the steps from the Tiglath-Pileser principle to Halpern’s conclusions. This framework builds on the theory developed by Rubinstein and Glazer in a series of papers [8, 9, 10] in which several types of situations are analysed in which one or more agents (the Speakers) wish to persuade another agent, the Listener (who plays the principal’s role),2^2 to make a decision according to their desiderata, e.g. to accept a particular request, to award a prize to them, or to believe a certain claim they make to be true. The Listener’s decision depends on what they conclude about a certain state of affairs, which can either concern the (particular) Speaker (e.g. whether the Speaker satisfies certain requirements) or some other facts about the world. Each Speaker knows the true state of affairs, whereas the Listener knows only the set SS of the possible states of affairs (and possibly some information about the Speaker’s goals and strategies). Each Speaker wishes to persuade the Listener regardless of the true state of affairs. Hence, the Speakers can lie, and will do so whenever declaring the true state of affairs would not favour them. However, the key feature of the scenarios in [8, 10] is that the Speakers’ lies are constrained by a particularly strong bounded rationality assumption that anchors their lies to the truth. Hence, in certain cases, the Listener can engineer the enquiry in such a way that they can approximate the true state of affairs to varying degrees of accuracy. To formally represent this idea, Rubinstein and Glazer introduce the notion of codex, a finite set LL of constraints, expressed as formulas of classical propositional logic, which the Speaker’s declared state is expected to satisfy. The specific way in which the Speakers’ lying ability is constrained is formalized by the guidance relation →L⊆S×S\to_{L} \subseteq S \times S associated with LL, where s→Ls′s \to_{L} s' if, whenever ss is the true state, the Speaker could declare s′s' in an attempt to satisfy codex LL.

Guidance relations are perspicuous formal representations of the Tiglath-Pileser principle in action: by knowing the Speaker’s guidance relation, the Listener is able to conclude that the true state must belong to the preimage of s′s' under →L\to_{L}, whenever the Speaker declares s′s' as the ‘true state’. Depending on the properties of the guidance relation, the Listener is able to gauge the information needed for their decision.

We introduce a class of relations, referred to as Tiglath-Pileser relations, which generalize guidance relations, and need not arise from codices. We pursue an axiomatic development of the properties of Tiglath-Pileser relations required for the successful application of the Tiglath-Pileser principle. We introduce a logic-based framework for the analysis of scenarios in which the Listener’s approximations of the Speaker’s Tiglath-Pileser relation are dynamically formed over the course of extended interaction, and apply it to two case studies.

The framework introduced in the present paper differs from that of [8, 9, 10] in important respects. Firstly, while in [8, 10], the constraints on the Speakers’ ability to lie are purely a result of their bounded rationality, we explore situations in which the background knowledge of the Listener imposes further constraints. Secondly, we are interested in exploring more open-ended scenarios than those considered in [8, 9, 10], where the latter mostly concern the structuring of selection procedures designed to elicit the truth in situations which can be modelled as one-off interactions, and the goals of the Speaker(s) are common knowledge among all agents. In contrast, we are interested in accounting for possibly extended interactions among agents, in which the goals of the Speaker(s) might not be explicitly known, and in fact, the Speaker(s) might have an interest in concealing them both from their primary intended audience, and from any agent who, playing the “historian’s role”, tries to interpret the whole communication. Thirdly, in contrast to [8, 9, 10], we take a dynamic perspective on how guidance relations arise in the Listener’s mind, and how this process can be manipulated by strategic Speakers to feed the Listener a false relation.3 Fourthly, in [8, 10], the Speakers are assumed to “not think strategically, i.e. not [to] take into consideration the other party’s actions”, but rather act as “problem solvers”, e.g. by (only) making sure that their declarations verify the codices they have been assigned; in contrast, we structure the analysis of the present paper according to a hierarchy of levels of strategic/cognitive [15, 13, 4] of the agents involved, the most basic level of which is that of the “non-strategically thinking” agents considered in [8, 10]. While, like [8, 10], we avoid an explicit game-theoretic treatment, this approach sets the stage for a game-theoretic development of signalling games with uncertain information and players acting at different levels of the strategic/cognitive hierarchy. In this perspective, this paper contributes novel insights about the formal modelling of trust based on inferring the motivations and persuasion strategy of other agents, rather than purely on their track-record of truth-telling in previous rounds of a game [12, 2].

The framework developed here has direct applications to detecting and managing disinformation. In adversarial information environments, strategic actors rarely invent entirely false narratives; instead, they often exploit, selectively amplify, and distort fragments of existing information to create misleading impressions [16]. By modelling propagandists as strategic Speakers constrained by the information their audience possess, our framework enables analysts, algorithms, and fact-checkers (Listeners) to attain a reliable approximation of the truth embedded within manipulated narratives. The present treatment of extended interactions and higher-order reasoning also models how sophisticated campaigns shape perceptions of both events and source motives.

Another area of application concerns AI safety and the governance of agentic systems. As AI agents become more capable, risks such as sycophancy and deceptive alignment become increasingly important [1]. This framework makes it possible to model AI systems as strategic Speakers interacting with human Listeners, and hence it can be used by supervisors to audit, probe, and constrain potentially deceptive behaviour. Such an approach aligns with emerging work on scalable oversight, interpretability, and evaluations designed to detect undesirable behaviours in advanced AI systems [3, 5].

Structure of the paper. In § 2, we give a detailed summary and analysis of the formalisms introduced in [8, 10]; in § 3, we further develop and refine this formal framework, with a particular focus on the relationship between codices and guidance relations; in § 4, we recast this formalism in a purely set-theoretic way; in this formal environment, the analysis is structured along a hierarchy of levels of strategic/cognitive awareness of the agents involved, successive levels of which correspond to deeper levels of sophistication in the agents’ reasoning about other agents’ motivations and Tiglath-Pileser relations; in § 5, we introduce a logic-based framework which accounts for how codices and their associated Tiglath-Pileser relations are dynamically generated in the mind of the Listener in the context of a narration which unfolds over several episodes, how these formal structures govern the process of formation of reliable beliefs in the Listener’s mind, how this process can be manipulated by strategic Speakers, and how the Listener can out-manoeuvre the Speaker in these manipulative attempts. This framework is applied to case studies in § 6 and § 7, the latter section collecting the formal account of Baruch Halpern’s analysis of King David’s story. We conclude in § 8.

Preliminaries

In this section, we review the formalisms introduced in [8, 10], where they are embedded in extensive motivating case studies of persuasion scenarios. Our presentation abstracts from the latter, while highlighting and developing the mathematical and formal aspects, preparing the ground for the introduction of the generalization to Tiglath-Pileser relations in § 3.

Basic formalism of `A model of persuasion'

For any (finite) set Prop={v1,…,vK}\mathrm{Prop} = \{v_1, \ldots, v_K\} of propositional variables4 with K≥2K \ge2, a profile, or state5, is a map s:Prop→2={0,1}s : \mathrm{Prop} \to2 = \{0,1\}; we let SS denote the set of states and write s⊨vs \models v iff s(v)=1s(v) = 1.

In [8], leader-follower scenarios are analysed in which a Listener announces and commits to a codex LL, which is a logically consistent set of propositional formulas in the language of VV that the profile presented by a Speaker must satisfy in order for the Speaker’s request to be accepted. Then, the Speaker (who knows the true profile) announces a profile that may or may not be the true one.

Each formula (rule) φ\varphi in the codex is required to have the following syntactic shape:

⋀y∈Yφy→φx\bigwedge_{y \in Y} \varphi_y \to\varphi_x

where YY is a nonempty subset of Prop\mathrm{Prop}, x∈Prop∖Yx \in\mathrm{Prop} \setminus Y, and each φv\varphi_v is either vv or ¬v\neg v (the negation of vv) for any v∈Propv \in\mathrm{Prop}. For example, the proposition v4∧¬v1→v3v_4 \land\neg v_1 \to v_3 can be a rule in a codex, but v1→¬v1v_1 \to\neg v_1 cannot.

For every formula φ\varphi and every set of formulas Γ\Gamma, let Var⁡(φ)\operatorname{Var}(\varphi) denote the set of variables occurring in φ\varphi and let Var⁡(Γ)=⋃φ∈ΓVar⁡(φ)\operatorname{Var}(\Gamma) = \bigcup_{\varphi\in\Gamma} \operatorname{Var}(\varphi). A rule φ\varphi is complete if Var⁡(φ)=Prop\operatorname{Var}(\varphi) = \mathrm{Prop}.

For any rule φ=⋀y∈Yφy→φx\varphi= \bigwedge_{y \in Y} \varphi_y \to\varphi_x, we denote its antecedent and consequent as follows:

a(φ)=⋀y∈Yφyandz(φ)=φx.a(\varphi) = \bigwedge_{y \in Y} \varphi_y \qquad\text{and} \qquad z(\varphi) = \varphi_x.

Codices are required to be not only satisfiable but also coherent, i.e. for all φ,ψ∈L\varphi, \psi\in L, if a(φ)a(\varphi) and a(ψ)a(\psi) are jointly satisfiable, then so are z(φ)z(\varphi) and z(ψ)z(\psi).

Definition 2.1 (Guidance relation). Let S=2KS = 2^K. The guidance relation →L\to_L associated with any codex LL is a binary relation →L⊆S×S\to_L \subseteq S \times S defined as follows: s→Ls′s \to_L s' iff for every variable v∈Propv \in\mathrm{Prop} such that s′(v)≠s(v)s'(v) \ne s(v), a rule φ∈L\varphi\in L exists such that:

  1. z(φ)∈{v,¬v}z(\varphi) \in\{v, \neg v\};

  1. s⊨a(φ)s \models a(\varphi) and s′⊨a(φ)s' \models a(\varphi), and

  1. s′⊨z(φ)s' \models z(\varphi).

Hence, if s→Ls′s \to_L s', then there exists at least one rule φ∈L\varphi\in L s.t. s⊭φs \not\models\varphi and s′⊨φs' \models\varphi, where the latter satisfaction is non-trivial, since s′⊨a(φ)s' \models a(\varphi). Codices LL consisting only of complete rules force their associated guidance relations to be subsets of the neighbourhood relation N⊆S×SN \subseteq S \times S s.t. sNs′sNs' iff some x∈Vx \in V exists s.t. s(v)=s′(v)s(v) = s'(v) for any v∈V∖{x}v \in V \setminus\{x\}. For any two such neighbors ss and s′s', consider the following complete rule

φ(s,s′):=⋀v∈Prop⁡∖{x}φv→φx.(1)\varphi(s,s') := \bigwedge_{v \in\operatorname{Prop} \setminus\{x\}} \varphi_v \to\varphi_x. \tag*{(1)}

Thus by construction, ss is the only profile which falsifies φ(s,s′)\varphi(s,s'), since every other profile would either verify its conclusion or falsify its premises, and s′s' is the unique profile which nontrivially satisfies φ(s,s′)\varphi(s,s'), since s′⊨a(φ(s,s′))s' \models a(\varphi(s,s')). Moreover, s→Ls′s \to_L s' for any codex LL that contains φ(s,s′)\varphi(s,s'). For any codex LL, let

T(L):={s∈S∣s⊨φ for all φ∈L}andP(L)={s∈S∣s→Ls′ for some s′∈T(L)}.T(L) := \{s \in S \mid s \models\varphi\text{ for all } \varphi\in L\} \quad\text{and}\quad P(L) = \{s \in S \mid s \to_L s' \text{ for some } s' \in T(L)\}.

That is, T(L)T(L) is the set of profiles that, whenever declared, the Listener would accept, and P(L)P(L) is the set of profiles from which the Speaker can be ‘guided’ by LL to some state which can persuade the Listener. By definition, T(L)⊆P(L)T(L) \subseteq P(L).

For any set AA of states, AA is implementable if A=P(L)A = P(L) for some codex LL. The set AA is truthfully implementable if A=T(L)=P(L)A = T(L) = P(L) for some codex LL.

Let T(→L)={s∈S∣∀t(s→Lt⇒s=t)}T(\to_L) = \{s \in S \mid\forall t(s \to_L t \Rightarrow s = t)\}.

Lemma 2.2 ([8], Lemma 2). For any codex LL, T(L)=T(→L)T(L) = T(\to_L).

Proof. If s∉T(L)s \notin T(L), then s⊭φs \not\models\varphi for some φ∈L\varphi\in L; that is, s⊨a(φ)s \models a(\varphi) and s⊭z(φ)=φxs \not\models z(\varphi) = \varphi_x for some x∈Prop⁡x \in\operatorname{Prop}. Thus, s→Ls′s \to_L s' where s′s' is the profile that differs from ss only in the truth value of the variable xx, which shows that s∉T(→L)s \notin T(\to_L). Conversely, let s∉T(→L)s \notin T(\to_L), i.e. s→Lts \to_L t for some profile t≠st \ne s. By definition of →L\to_L, this implies that s⊭φs \not\models\varphi for some φ∈L\varphi\in L; therefore, s∉T(L)s \notin T(L). □

Definition 2.3 (Canonical and quasi-canonical codices). A codex LL is canonical if:

  1. LL consists of complete rules.

  2. For every ss, there is at most one t≠st \ne s such that s→Lts \to_L t.

  3. For every s∈P(L)∖T(L)s \in P(L) \setminus T(L), there is some r∈S∖P(L)r \in S \setminus P(L) such that r→Lsr \to_L s.

A codex LL is quasi-canonical if it satisfies the conditions in the first two items above.6

Proposition 2.4 ([8], Proposition 1). If A⊆SA \subseteq S is truthfully implementable, then it is truthfully implementable by a canonical codex.

We conclude this subsection by reporting on some examples discussed in [8]. Let Prop⁡={v1,v2,v3}\operatorname{Prop}=\{v_1,v_2,v_3\}. The codex

References

  1. [1]D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mané, Concrete problems in AI safety, arXiv preprint arXiv:1606.06565, (2016).arxiv.org/abs/1606.06565
  2. [2]A. E. Atakan and M. Ekmekci, Reputation in long-run relationships, The Review of Economic Studies, 79 (2012), pp. 451–480.DOI
  3. [3]Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, C. Chen, C. Olsson, C. Olah, D. Hernandez, D. Drain, D. Ganguli, D. Li, E. Tran-Johnson, E. Perez, J. Kerr, J. Mueller, J. Ladish, J. Landau, K. Ndousse, K. Lukosuite, L. Lovitt, M. Sellitto, N. Elhage, N. Schiefer, N. Mercado, N. DasSarma, R. Lasenby, R. Larson, S. Ringer, S. Johnston, S. Kravec, S. Sheer El Showk, S. Fort, T. Lanham, T. Telleen-Lawton, T. Conerly, T. Henighan, T. Hume, S. R. Bowman, Z. Hatfield-Dodds, B. Mann, D. Amodei, N. Joseph, S. McCandlish, T. Brown, and J. Kaplan, Constitutional AI: Harmlessness from AI feedback, arXiv preprint arXiv:2212.08073, (2022).arxiv.org/abs/2212.08073
  4. [4]C. F. Camerer, T.-H. Ho, and J.-K. Chong, A cognitive hierarchy model of games, The quarterly journal of economics, 119 (2004), pp. 861–898.DOI
  5. [5]P. F. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei, Deep reinforcement learning from human preferences, Advances in neural information processing systems, 30 (2017).arxiv.org/abs/1706.03741
  6. [6]W. Conradie, S. Frittella, A. Palmigiano, A. Tzimoulis, and N. Wijnberg, Probabilistic epistemic updates on algebras, ACM Transactions on Computational Logic (TOCL), 20 (2019), pp. 1–75.
  7. [7]R. Fagin, J. Y. Halpern, Y. Moses, and M. Vardi, Reasoning about knowledge, 2004.
  8. [8]J. Glazer and A. Rubinstein, A model of persuasion with boundedly rational agents, Journal of Political Economy, 120 (2012), pp. 1057–1082.
  9. [9]——, Complex questionnaires, Econometrica, 82 (2014), pp. 1529–1541.
  10. [10]——, Magical implementation, (2024).
  11. [11]B. Halpern, David’s secret demons: messiah, murderer, traitor, king, Wm. B. Eerdmans Publishing, 2003.
  12. [12]D. M. Kreps and R. Wilson, Reputation and imperfect information, Journal of Economic Theory, 27 (1982), pp. 253–279.DOI
  13. [13]R. Nagel, Unraveling in guessing games: An experimental study, The American economic review, 85 (1995), pp. 1313–1326.
  14. [14]S. A. Ross, The economic theory of agency: The principal’s problem, The American economic review, 63 (1973), pp. 134–139.
  15. [15]D. O. Stahl II and P. W. Wilson, Experimental evidence on players’ models of other players, Journal of economic behavior & organization, 25 (1994), pp. 309–327.DOI
  16. [16]C. Wardle and H. Derakhshan, Information disorder: Toward an interdisciplinary framework for research and policymaking, vol. 27, Council of Europe Strasbourg, 2017.

Paper details

vlnc:W_ukgap94u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents